Commit | Line | Data |
---|---|---|
1439f94c | 1 | // SPDX-License-Identifier: GPL-2.0-only |
6a46079c AK |
2 | /* |
3 | * Copyright (C) 2008, 2009 Intel Corporation | |
4 | * Authors: Andi Kleen, Fengguang Wu | |
5 | * | |
6a46079c | 6 | * High level machine check handler. Handles pages reported by the |
1c80b990 | 7 | * hardware as being corrupted usually due to a multi-bit ECC memory or cache |
6a46079c | 8 | * failure. |
1c80b990 AK |
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. | |
6a46079c AK |
12 | * |
13 | * Handles page cache pages in various states. The tricky part | |
1c80b990 AK |
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. | |
e0de78df AK |
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 | |
799fb82a | 27 | * tools/mm/page-types when running a real workload. |
1c80b990 AK |
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. | |
6a46079c | 35 | */ |
96f96763 KW |
36 | |
37 | #define pr_fmt(fmt) "Memory failure: " fmt | |
38 | ||
6a46079c AK |
39 | #include <linux/kernel.h> |
40 | #include <linux/mm.h> | |
41 | #include <linux/page-flags.h> | |
478c5ffc | 42 | #include <linux/kernel-page-flags.h> |
3f07c014 | 43 | #include <linux/sched/signal.h> |
29930025 | 44 | #include <linux/sched/task.h> |
96c84dde | 45 | #include <linux/dax.h> |
01e00f88 | 46 | #include <linux/ksm.h> |
6a46079c | 47 | #include <linux/rmap.h> |
b9e15baf | 48 | #include <linux/export.h> |
6a46079c AK |
49 | #include <linux/pagemap.h> |
50 | #include <linux/swap.h> | |
51 | #include <linux/backing-dev.h> | |
facb6011 | 52 | #include <linux/migrate.h> |
facb6011 | 53 | #include <linux/suspend.h> |
5a0e3ad6 | 54 | #include <linux/slab.h> |
bf998156 | 55 | #include <linux/swapops.h> |
7af446a8 | 56 | #include <linux/hugetlb.h> |
20d6c96b | 57 | #include <linux/memory_hotplug.h> |
5db8a73a | 58 | #include <linux/mm_inline.h> |
6100e34b | 59 | #include <linux/memremap.h> |
ea8f5fb8 | 60 | #include <linux/kfifo.h> |
a5f65109 | 61 | #include <linux/ratelimit.h> |
d4ae9916 | 62 | #include <linux/page-isolation.h> |
a3f5d80e | 63 | #include <linux/pagewalk.h> |
a7605426 | 64 | #include <linux/shmem_fs.h> |
014bb1de | 65 | #include "swap.h" |
6a46079c | 66 | #include "internal.h" |
97f0b134 | 67 | #include "ras/ras_event.h" |
6a46079c AK |
68 | |
69 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
70 | ||
71 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
72 | ||
293c07e3 | 73 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 74 | |
67f22ba7 | 75 | static bool hw_memory_failure __read_mostly = false; |
76 | ||
a46c9304 | 77 | inline void num_poisoned_pages_inc(unsigned long pfn) |
d027122d NH |
78 | { |
79 | atomic_long_inc(&num_poisoned_pages); | |
5033091d | 80 | memblk_nr_poison_inc(pfn); |
d027122d NH |
81 | } |
82 | ||
5033091d | 83 | inline void num_poisoned_pages_sub(unsigned long pfn, long i) |
d027122d NH |
84 | { |
85 | atomic_long_sub(i, &num_poisoned_pages); | |
5033091d NH |
86 | if (pfn != -1UL) |
87 | memblk_nr_poison_sub(pfn, i); | |
d027122d NH |
88 | } |
89 | ||
7453bf62 NH |
90 | /* |
91 | * Return values: | |
92 | * 1: the page is dissolved (if needed) and taken off from buddy, | |
93 | * 0: the page is dissolved (if needed) and not taken off from buddy, | |
94 | * < 0: failed to dissolve. | |
95 | */ | |
96 | static int __page_handle_poison(struct page *page) | |
510d25c9 | 97 | { |
f87060d3 | 98 | int ret; |
510d25c9 NH |
99 | |
100 | zone_pcp_disable(page_zone(page)); | |
101 | ret = dissolve_free_huge_page(page); | |
102 | if (!ret) | |
103 | ret = take_page_off_buddy(page); | |
104 | zone_pcp_enable(page_zone(page)); | |
105 | ||
7453bf62 | 106 | return ret; |
510d25c9 NH |
107 | } |
108 | ||
6b9a217e | 109 | static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release) |
06be6ff3 | 110 | { |
6b9a217e OS |
111 | if (hugepage_or_freepage) { |
112 | /* | |
113 | * Doing this check for free pages is also fine since dissolve_free_huge_page | |
114 | * returns 0 for non-hugetlb pages as well. | |
115 | */ | |
7453bf62 | 116 | if (__page_handle_poison(page) <= 0) |
6b9a217e OS |
117 | /* |
118 | * We could fail to take off the target page from buddy | |
f0953a1b | 119 | * for example due to racy page allocation, but that's |
6b9a217e OS |
120 | * acceptable because soft-offlined page is not broken |
121 | * and if someone really want to use it, they should | |
122 | * take it. | |
123 | */ | |
124 | return false; | |
125 | } | |
126 | ||
06be6ff3 | 127 | SetPageHWPoison(page); |
79f5f8fa OS |
128 | if (release) |
129 | put_page(page); | |
06be6ff3 | 130 | page_ref_inc(page); |
a46c9304 | 131 | num_poisoned_pages_inc(page_to_pfn(page)); |
6b9a217e OS |
132 | |
133 | return true; | |
06be6ff3 OS |
134 | } |
135 | ||
27df5068 AK |
136 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
137 | ||
1bfe5feb | 138 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
139 | u32 hwpoison_filter_dev_major = ~0U; |
140 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
141 | u64 hwpoison_filter_flags_mask; |
142 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 143 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
144 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
145 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
146 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
147 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
148 | |
149 | static int hwpoison_filter_dev(struct page *p) | |
150 | { | |
151 | struct address_space *mapping; | |
152 | dev_t dev; | |
153 | ||
154 | if (hwpoison_filter_dev_major == ~0U && | |
155 | hwpoison_filter_dev_minor == ~0U) | |
156 | return 0; | |
157 | ||
7c116f2b WF |
158 | mapping = page_mapping(p); |
159 | if (mapping == NULL || mapping->host == NULL) | |
160 | return -EINVAL; | |
161 | ||
162 | dev = mapping->host->i_sb->s_dev; | |
163 | if (hwpoison_filter_dev_major != ~0U && | |
164 | hwpoison_filter_dev_major != MAJOR(dev)) | |
165 | return -EINVAL; | |
166 | if (hwpoison_filter_dev_minor != ~0U && | |
167 | hwpoison_filter_dev_minor != MINOR(dev)) | |
168 | return -EINVAL; | |
169 | ||
170 | return 0; | |
171 | } | |
172 | ||
478c5ffc WF |
173 | static int hwpoison_filter_flags(struct page *p) |
174 | { | |
175 | if (!hwpoison_filter_flags_mask) | |
176 | return 0; | |
177 | ||
178 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
179 | hwpoison_filter_flags_value) | |
180 | return 0; | |
181 | else | |
182 | return -EINVAL; | |
183 | } | |
184 | ||
4fd466eb AK |
185 | /* |
186 | * This allows stress tests to limit test scope to a collection of tasks | |
187 | * by putting them under some memcg. This prevents killing unrelated/important | |
188 | * processes such as /sbin/init. Note that the target task may share clean | |
189 | * pages with init (eg. libc text), which is harmless. If the target task | |
190 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
191 | * is also included in the memcg. At last, due to race conditions this filter | |
192 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
193 | * a freed page. | |
194 | */ | |
94a59fb3 | 195 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
196 | u64 hwpoison_filter_memcg; |
197 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
198 | static int hwpoison_filter_task(struct page *p) | |
199 | { | |
4fd466eb AK |
200 | if (!hwpoison_filter_memcg) |
201 | return 0; | |
202 | ||
94a59fb3 | 203 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
204 | return -EINVAL; |
205 | ||
206 | return 0; | |
207 | } | |
208 | #else | |
209 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
210 | #endif | |
211 | ||
7c116f2b WF |
212 | int hwpoison_filter(struct page *p) |
213 | { | |
1bfe5feb HL |
214 | if (!hwpoison_filter_enable) |
215 | return 0; | |
216 | ||
7c116f2b WF |
217 | if (hwpoison_filter_dev(p)) |
218 | return -EINVAL; | |
219 | ||
478c5ffc WF |
220 | if (hwpoison_filter_flags(p)) |
221 | return -EINVAL; | |
222 | ||
4fd466eb AK |
223 | if (hwpoison_filter_task(p)) |
224 | return -EINVAL; | |
225 | ||
7c116f2b WF |
226 | return 0; |
227 | } | |
27df5068 AK |
228 | #else |
229 | int hwpoison_filter(struct page *p) | |
230 | { | |
231 | return 0; | |
232 | } | |
233 | #endif | |
234 | ||
7c116f2b WF |
235 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
236 | ||
ae1139ec DW |
237 | /* |
238 | * Kill all processes that have a poisoned page mapped and then isolate | |
239 | * the page. | |
240 | * | |
241 | * General strategy: | |
242 | * Find all processes having the page mapped and kill them. | |
243 | * But we keep a page reference around so that the page is not | |
244 | * actually freed yet. | |
245 | * Then stash the page away | |
246 | * | |
247 | * There's no convenient way to get back to mapped processes | |
248 | * from the VMAs. So do a brute-force search over all | |
249 | * running processes. | |
250 | * | |
251 | * Remember that machine checks are not common (or rather | |
252 | * if they are common you have other problems), so this shouldn't | |
253 | * be a performance issue. | |
254 | * | |
255 | * Also there are some races possible while we get from the | |
256 | * error detection to actually handle it. | |
257 | */ | |
258 | ||
259 | struct to_kill { | |
260 | struct list_head nd; | |
261 | struct task_struct *tsk; | |
262 | unsigned long addr; | |
263 | short size_shift; | |
ae1139ec DW |
264 | }; |
265 | ||
6a46079c | 266 | /* |
7329bbeb TL |
267 | * Send all the processes who have the page mapped a signal. |
268 | * ``action optional'' if they are not immediately affected by the error | |
269 | * ``action required'' if error happened in current execution context | |
6a46079c | 270 | */ |
ae1139ec | 271 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 272 | { |
ae1139ec DW |
273 | struct task_struct *t = tk->tsk; |
274 | short addr_lsb = tk->size_shift; | |
872e9a20 | 275 | int ret = 0; |
6a46079c | 276 | |
96f96763 | 277 | pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", |
872e9a20 | 278 | pfn, t->comm, t->pid); |
7329bbeb | 279 | |
49775047 ML |
280 | if ((flags & MF_ACTION_REQUIRED) && (t == current)) |
281 | ret = force_sig_mceerr(BUS_MCEERR_AR, | |
282 | (void __user *)tk->addr, addr_lsb); | |
283 | else | |
7329bbeb | 284 | /* |
49775047 ML |
285 | * Signal other processes sharing the page if they have |
286 | * PF_MCE_EARLY set. | |
7329bbeb TL |
287 | * Don't use force here, it's convenient if the signal |
288 | * can be temporarily blocked. | |
289 | * This could cause a loop when the user sets SIGBUS | |
290 | * to SIG_IGN, but hopefully no one will do that? | |
291 | */ | |
ae1139ec | 292 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
9cf28191 | 293 | addr_lsb, t); |
6a46079c | 294 | if (ret < 0) |
96f96763 | 295 | pr_info("Error sending signal to %s:%d: %d\n", |
1170532b | 296 | t->comm, t->pid, ret); |
6a46079c AK |
297 | return ret; |
298 | } | |
299 | ||
588f9ce6 | 300 | /* |
47e431f4 | 301 | * Unknown page type encountered. Try to check whether it can turn PageLRU by |
d0505e9f | 302 | * lru_add_drain_all. |
588f9ce6 | 303 | */ |
d0505e9f | 304 | void shake_page(struct page *p) |
588f9ce6 | 305 | { |
8bcb74de NH |
306 | if (PageHuge(p)) |
307 | return; | |
308 | ||
588f9ce6 AK |
309 | if (!PageSlab(p)) { |
310 | lru_add_drain_all(); | |
588f9ce6 AK |
311 | if (PageLRU(p) || is_free_buddy_page(p)) |
312 | return; | |
313 | } | |
facb6011 | 314 | |
588f9ce6 | 315 | /* |
d0505e9f YS |
316 | * TODO: Could shrink slab caches here if a lightweight range-based |
317 | * shrinker will be available. | |
588f9ce6 AK |
318 | */ |
319 | } | |
320 | EXPORT_SYMBOL_GPL(shake_page); | |
321 | ||
c36e2024 SR |
322 | static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct *vma, |
323 | unsigned long address) | |
6100e34b | 324 | { |
5c91c0e7 | 325 | unsigned long ret = 0; |
6100e34b DW |
326 | pgd_t *pgd; |
327 | p4d_t *p4d; | |
328 | pud_t *pud; | |
329 | pmd_t *pmd; | |
330 | pte_t *pte; | |
331 | ||
a994402b | 332 | VM_BUG_ON_VMA(address == -EFAULT, vma); |
6100e34b DW |
333 | pgd = pgd_offset(vma->vm_mm, address); |
334 | if (!pgd_present(*pgd)) | |
335 | return 0; | |
336 | p4d = p4d_offset(pgd, address); | |
337 | if (!p4d_present(*p4d)) | |
338 | return 0; | |
339 | pud = pud_offset(p4d, address); | |
340 | if (!pud_present(*pud)) | |
341 | return 0; | |
342 | if (pud_devmap(*pud)) | |
343 | return PUD_SHIFT; | |
344 | pmd = pmd_offset(pud, address); | |
345 | if (!pmd_present(*pmd)) | |
346 | return 0; | |
347 | if (pmd_devmap(*pmd)) | |
348 | return PMD_SHIFT; | |
349 | pte = pte_offset_map(pmd, address); | |
5c91c0e7 QZ |
350 | if (pte_present(*pte) && pte_devmap(*pte)) |
351 | ret = PAGE_SHIFT; | |
352 | pte_unmap(pte); | |
353 | return ret; | |
6100e34b | 354 | } |
6a46079c AK |
355 | |
356 | /* | |
357 | * Failure handling: if we can't find or can't kill a process there's | |
358 | * not much we can do. We just print a message and ignore otherwise. | |
359 | */ | |
360 | ||
ac87ca0e DW |
361 | #define FSDAX_INVALID_PGOFF ULONG_MAX |
362 | ||
6a46079c AK |
363 | /* |
364 | * Schedule a process for later kill. | |
365 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
c36e2024 | 366 | * |
ac87ca0e DW |
367 | * Note: @fsdax_pgoff is used only when @p is a fsdax page and a |
368 | * filesystem with a memory failure handler has claimed the | |
369 | * memory_failure event. In all other cases, page->index and | |
370 | * page->mapping are sufficient for mapping the page back to its | |
371 | * corresponding user virtual address. | |
6a46079c AK |
372 | */ |
373 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
c36e2024 SR |
374 | pgoff_t fsdax_pgoff, struct vm_area_struct *vma, |
375 | struct list_head *to_kill) | |
6a46079c AK |
376 | { |
377 | struct to_kill *tk; | |
378 | ||
996ff7a0 JC |
379 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
380 | if (!tk) { | |
96f96763 | 381 | pr_err("Out of memory while machine check handling\n"); |
996ff7a0 | 382 | return; |
6a46079c | 383 | } |
996ff7a0 | 384 | |
6a46079c | 385 | tk->addr = page_address_in_vma(p, vma); |
c36e2024 | 386 | if (is_zone_device_page(p)) { |
ac87ca0e | 387 | if (fsdax_pgoff != FSDAX_INVALID_PGOFF) |
c36e2024 SR |
388 | tk->addr = vma_pgoff_address(fsdax_pgoff, 1, vma); |
389 | tk->size_shift = dev_pagemap_mapping_shift(vma, tk->addr); | |
390 | } else | |
75068518 | 391 | tk->size_shift = page_shift(compound_head(p)); |
6a46079c AK |
392 | |
393 | /* | |
3d7fed4a JC |
394 | * Send SIGKILL if "tk->addr == -EFAULT". Also, as |
395 | * "tk->size_shift" is always non-zero for !is_zone_device_page(), | |
396 | * so "tk->size_shift == 0" effectively checks no mapping on | |
397 | * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times | |
398 | * to a process' address space, it's possible not all N VMAs | |
399 | * contain mappings for the page, but at least one VMA does. | |
400 | * Only deliver SIGBUS with payload derived from the VMA that | |
401 | * has a mapping for the page. | |
6a46079c | 402 | */ |
3d7fed4a | 403 | if (tk->addr == -EFAULT) { |
96f96763 | 404 | pr_info("Unable to find user space address %lx in %s\n", |
6a46079c | 405 | page_to_pfn(p), tsk->comm); |
3d7fed4a JC |
406 | } else if (tk->size_shift == 0) { |
407 | kfree(tk); | |
408 | return; | |
6a46079c | 409 | } |
996ff7a0 | 410 | |
6a46079c AK |
411 | get_task_struct(tsk); |
412 | tk->tsk = tsk; | |
413 | list_add_tail(&tk->nd, to_kill); | |
414 | } | |
415 | ||
416 | /* | |
417 | * Kill the processes that have been collected earlier. | |
418 | * | |
a21c184f ML |
419 | * Only do anything when FORCEKILL is set, otherwise just free the |
420 | * list (this is used for clean pages which do not need killing) | |
6a46079c AK |
421 | * Also when FAIL is set do a force kill because something went |
422 | * wrong earlier. | |
423 | */ | |
ae1139ec DW |
424 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
425 | unsigned long pfn, int flags) | |
6a46079c AK |
426 | { |
427 | struct to_kill *tk, *next; | |
428 | ||
54f9555d | 429 | list_for_each_entry_safe(tk, next, to_kill, nd) { |
6751ed65 | 430 | if (forcekill) { |
6a46079c | 431 | /* |
af901ca1 | 432 | * In case something went wrong with munmapping |
6a46079c AK |
433 | * make sure the process doesn't catch the |
434 | * signal and then access the memory. Just kill it. | |
6a46079c | 435 | */ |
3d7fed4a | 436 | if (fail || tk->addr == -EFAULT) { |
96f96763 | 437 | pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 438 | pfn, tk->tsk->comm, tk->tsk->pid); |
6376360e NH |
439 | do_send_sig_info(SIGKILL, SEND_SIG_PRIV, |
440 | tk->tsk, PIDTYPE_PID); | |
6a46079c AK |
441 | } |
442 | ||
443 | /* | |
444 | * In theory the process could have mapped | |
445 | * something else on the address in-between. We could | |
446 | * check for that, but we need to tell the | |
447 | * process anyways. | |
448 | */ | |
ae1139ec | 449 | else if (kill_proc(tk, pfn, flags) < 0) |
96f96763 | 450 | pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 451 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c | 452 | } |
54f9555d | 453 | list_del(&tk->nd); |
6a46079c AK |
454 | put_task_struct(tk->tsk); |
455 | kfree(tk); | |
456 | } | |
457 | } | |
458 | ||
3ba08129 NH |
459 | /* |
460 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
461 | * on behalf of the thread group. Return task_struct of the (first found) | |
462 | * dedicated thread if found, and return NULL otherwise. | |
463 | * | |
464 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
465 | * have to call rcu_read_lock/unlock() in this function. | |
466 | */ | |
467 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 468 | { |
3ba08129 NH |
469 | struct task_struct *t; |
470 | ||
4e018b45 NH |
471 | for_each_thread(tsk, t) { |
472 | if (t->flags & PF_MCE_PROCESS) { | |
473 | if (t->flags & PF_MCE_EARLY) | |
474 | return t; | |
475 | } else { | |
476 | if (sysctl_memory_failure_early_kill) | |
477 | return t; | |
478 | } | |
479 | } | |
3ba08129 NH |
480 | return NULL; |
481 | } | |
482 | ||
483 | /* | |
484 | * Determine whether a given process is "early kill" process which expects | |
485 | * to be signaled when some page under the process is hwpoisoned. | |
486 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
30c9cf49 | 487 | * specified) if the process is "early kill" and otherwise returns NULL. |
03151c6e | 488 | * |
30c9cf49 AY |
489 | * Note that the above is true for Action Optional case. For Action Required |
490 | * case, it's only meaningful to the current thread which need to be signaled | |
491 | * with SIGBUS, this error is Action Optional for other non current | |
492 | * processes sharing the same error page,if the process is "early kill", the | |
493 | * task_struct of the dedicated thread will also be returned. | |
3ba08129 NH |
494 | */ |
495 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
496 | int force_early) | |
497 | { | |
6a46079c | 498 | if (!tsk->mm) |
3ba08129 | 499 | return NULL; |
30c9cf49 AY |
500 | /* |
501 | * Comparing ->mm here because current task might represent | |
502 | * a subthread, while tsk always points to the main thread. | |
503 | */ | |
504 | if (force_early && tsk->mm == current->mm) | |
505 | return current; | |
506 | ||
4e018b45 | 507 | return find_early_kill_thread(tsk); |
6a46079c AK |
508 | } |
509 | ||
510 | /* | |
511 | * Collect processes when the error hit an anonymous page. | |
512 | */ | |
513 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 514 | int force_early) |
6a46079c | 515 | { |
9595d769 | 516 | struct folio *folio = page_folio(page); |
6a46079c AK |
517 | struct vm_area_struct *vma; |
518 | struct task_struct *tsk; | |
519 | struct anon_vma *av; | |
bf181b9f | 520 | pgoff_t pgoff; |
6a46079c | 521 | |
6d4675e6 | 522 | av = folio_lock_anon_vma_read(folio, NULL); |
6a46079c | 523 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
524 | return; |
525 | ||
a0f7a756 | 526 | pgoff = page_to_pgoff(page); |
9b679320 | 527 | read_lock(&tasklist_lock); |
6a46079c | 528 | for_each_process (tsk) { |
5beb4930 | 529 | struct anon_vma_chain *vmac; |
3ba08129 | 530 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 531 | |
3ba08129 | 532 | if (!t) |
6a46079c | 533 | continue; |
bf181b9f ML |
534 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
535 | pgoff, pgoff) { | |
5beb4930 | 536 | vma = vmac->vma; |
36537a67 ML |
537 | if (vma->vm_mm != t->mm) |
538 | continue; | |
6a46079c AK |
539 | if (!page_mapped_in_vma(page, vma)) |
540 | continue; | |
36537a67 | 541 | add_to_kill(t, page, FSDAX_INVALID_PGOFF, vma, to_kill); |
6a46079c AK |
542 | } |
543 | } | |
6a46079c | 544 | read_unlock(&tasklist_lock); |
0c826c0b | 545 | anon_vma_unlock_read(av); |
6a46079c AK |
546 | } |
547 | ||
548 | /* | |
549 | * Collect processes when the error hit a file mapped page. | |
550 | */ | |
551 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 552 | int force_early) |
6a46079c AK |
553 | { |
554 | struct vm_area_struct *vma; | |
555 | struct task_struct *tsk; | |
6a46079c | 556 | struct address_space *mapping = page->mapping; |
c43bc03d | 557 | pgoff_t pgoff; |
6a46079c | 558 | |
d28eb9c8 | 559 | i_mmap_lock_read(mapping); |
9b679320 | 560 | read_lock(&tasklist_lock); |
c43bc03d | 561 | pgoff = page_to_pgoff(page); |
6a46079c | 562 | for_each_process(tsk) { |
3ba08129 | 563 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 564 | |
3ba08129 | 565 | if (!t) |
6a46079c | 566 | continue; |
6b2dbba8 | 567 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
568 | pgoff) { |
569 | /* | |
570 | * Send early kill signal to tasks where a vma covers | |
571 | * the page but the corrupted page is not necessarily | |
572 | * mapped it in its pte. | |
573 | * Assume applications who requested early kill want | |
574 | * to be informed of all such data corruptions. | |
575 | */ | |
3ba08129 | 576 | if (vma->vm_mm == t->mm) |
ac87ca0e DW |
577 | add_to_kill(t, page, FSDAX_INVALID_PGOFF, vma, |
578 | to_kill); | |
6a46079c AK |
579 | } |
580 | } | |
6a46079c | 581 | read_unlock(&tasklist_lock); |
d28eb9c8 | 582 | i_mmap_unlock_read(mapping); |
6a46079c AK |
583 | } |
584 | ||
c36e2024 SR |
585 | #ifdef CONFIG_FS_DAX |
586 | /* | |
587 | * Collect processes when the error hit a fsdax page. | |
588 | */ | |
589 | static void collect_procs_fsdax(struct page *page, | |
590 | struct address_space *mapping, pgoff_t pgoff, | |
591 | struct list_head *to_kill) | |
592 | { | |
593 | struct vm_area_struct *vma; | |
594 | struct task_struct *tsk; | |
595 | ||
596 | i_mmap_lock_read(mapping); | |
597 | read_lock(&tasklist_lock); | |
598 | for_each_process(tsk) { | |
599 | struct task_struct *t = task_early_kill(tsk, true); | |
600 | ||
601 | if (!t) | |
602 | continue; | |
603 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { | |
604 | if (vma->vm_mm == t->mm) | |
605 | add_to_kill(t, page, pgoff, vma, to_kill); | |
606 | } | |
607 | } | |
608 | read_unlock(&tasklist_lock); | |
609 | i_mmap_unlock_read(mapping); | |
610 | } | |
611 | #endif /* CONFIG_FS_DAX */ | |
612 | ||
6a46079c AK |
613 | /* |
614 | * Collect the processes who have the corrupted page mapped to kill. | |
6a46079c | 615 | */ |
74614de1 TL |
616 | static void collect_procs(struct page *page, struct list_head *tokill, |
617 | int force_early) | |
6a46079c | 618 | { |
6a46079c AK |
619 | if (!page->mapping) |
620 | return; | |
621 | ||
6a46079c | 622 | if (PageAnon(page)) |
996ff7a0 | 623 | collect_procs_anon(page, tokill, force_early); |
6a46079c | 624 | else |
996ff7a0 | 625 | collect_procs_file(page, tokill, force_early); |
6a46079c AK |
626 | } |
627 | ||
a3f5d80e NH |
628 | struct hwp_walk { |
629 | struct to_kill tk; | |
630 | unsigned long pfn; | |
631 | int flags; | |
632 | }; | |
633 | ||
634 | static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift) | |
635 | { | |
636 | tk->addr = addr; | |
637 | tk->size_shift = shift; | |
638 | } | |
639 | ||
640 | static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift, | |
641 | unsigned long poisoned_pfn, struct to_kill *tk) | |
642 | { | |
643 | unsigned long pfn = 0; | |
644 | ||
645 | if (pte_present(pte)) { | |
646 | pfn = pte_pfn(pte); | |
647 | } else { | |
648 | swp_entry_t swp = pte_to_swp_entry(pte); | |
649 | ||
650 | if (is_hwpoison_entry(swp)) | |
0d206b5d | 651 | pfn = swp_offset_pfn(swp); |
a3f5d80e NH |
652 | } |
653 | ||
654 | if (!pfn || pfn != poisoned_pfn) | |
655 | return 0; | |
656 | ||
657 | set_to_kill(tk, addr, shift); | |
658 | return 1; | |
659 | } | |
660 | ||
661 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
662 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
663 | struct hwp_walk *hwp) | |
664 | { | |
665 | pmd_t pmd = *pmdp; | |
666 | unsigned long pfn; | |
667 | unsigned long hwpoison_vaddr; | |
668 | ||
669 | if (!pmd_present(pmd)) | |
670 | return 0; | |
671 | pfn = pmd_pfn(pmd); | |
672 | if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) { | |
673 | hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT); | |
674 | set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT); | |
675 | return 1; | |
676 | } | |
677 | return 0; | |
678 | } | |
679 | #else | |
680 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
681 | struct hwp_walk *hwp) | |
682 | { | |
683 | return 0; | |
684 | } | |
685 | #endif | |
686 | ||
687 | static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr, | |
688 | unsigned long end, struct mm_walk *walk) | |
689 | { | |
f142e707 | 690 | struct hwp_walk *hwp = walk->private; |
a3f5d80e | 691 | int ret = 0; |
ea3732f7 | 692 | pte_t *ptep, *mapped_pte; |
a3f5d80e NH |
693 | spinlock_t *ptl; |
694 | ||
695 | ptl = pmd_trans_huge_lock(pmdp, walk->vma); | |
696 | if (ptl) { | |
697 | ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp); | |
698 | spin_unlock(ptl); | |
699 | goto out; | |
700 | } | |
701 | ||
702 | if (pmd_trans_unstable(pmdp)) | |
703 | goto out; | |
704 | ||
ea3732f7 ML |
705 | mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, |
706 | addr, &ptl); | |
a3f5d80e NH |
707 | for (; addr != end; ptep++, addr += PAGE_SIZE) { |
708 | ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT, | |
709 | hwp->pfn, &hwp->tk); | |
710 | if (ret == 1) | |
711 | break; | |
712 | } | |
ea3732f7 | 713 | pte_unmap_unlock(mapped_pte, ptl); |
a3f5d80e NH |
714 | out: |
715 | cond_resched(); | |
716 | return ret; | |
717 | } | |
718 | ||
719 | #ifdef CONFIG_HUGETLB_PAGE | |
720 | static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask, | |
721 | unsigned long addr, unsigned long end, | |
722 | struct mm_walk *walk) | |
723 | { | |
f142e707 | 724 | struct hwp_walk *hwp = walk->private; |
a3f5d80e NH |
725 | pte_t pte = huge_ptep_get(ptep); |
726 | struct hstate *h = hstate_vma(walk->vma); | |
727 | ||
728 | return check_hwpoisoned_entry(pte, addr, huge_page_shift(h), | |
729 | hwp->pfn, &hwp->tk); | |
730 | } | |
731 | #else | |
732 | #define hwpoison_hugetlb_range NULL | |
733 | #endif | |
734 | ||
ba9eb3ce | 735 | static const struct mm_walk_ops hwp_walk_ops = { |
a3f5d80e NH |
736 | .pmd_entry = hwpoison_pte_range, |
737 | .hugetlb_entry = hwpoison_hugetlb_range, | |
738 | }; | |
739 | ||
740 | /* | |
741 | * Sends SIGBUS to the current process with error info. | |
742 | * | |
743 | * This function is intended to handle "Action Required" MCEs on already | |
744 | * hardware poisoned pages. They could happen, for example, when | |
745 | * memory_failure() failed to unmap the error page at the first call, or | |
746 | * when multiple local machine checks happened on different CPUs. | |
747 | * | |
748 | * MCE handler currently has no easy access to the error virtual address, | |
749 | * so this function walks page table to find it. The returned virtual address | |
750 | * is proper in most cases, but it could be wrong when the application | |
751 | * process has multiple entries mapping the error page. | |
752 | */ | |
753 | static int kill_accessing_process(struct task_struct *p, unsigned long pfn, | |
754 | int flags) | |
755 | { | |
756 | int ret; | |
757 | struct hwp_walk priv = { | |
758 | .pfn = pfn, | |
759 | }; | |
760 | priv.tk.tsk = p; | |
761 | ||
77677cdb SX |
762 | if (!p->mm) |
763 | return -EFAULT; | |
764 | ||
a3f5d80e NH |
765 | mmap_read_lock(p->mm); |
766 | ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops, | |
767 | (void *)&priv); | |
768 | if (ret == 1 && priv.tk.addr) | |
769 | kill_proc(&priv.tk, pfn, flags); | |
046545a6 NH |
770 | else |
771 | ret = 0; | |
a3f5d80e | 772 | mmap_read_unlock(p->mm); |
046545a6 | 773 | return ret > 0 ? -EHWPOISON : -EFAULT; |
a3f5d80e NH |
774 | } |
775 | ||
6a46079c | 776 | static const char *action_name[] = { |
cc637b17 XX |
777 | [MF_IGNORED] = "Ignored", |
778 | [MF_FAILED] = "Failed", | |
779 | [MF_DELAYED] = "Delayed", | |
780 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
781 | }; |
782 | ||
783 | static const char * const action_page_types[] = { | |
cc637b17 XX |
784 | [MF_MSG_KERNEL] = "reserved kernel page", |
785 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
786 | [MF_MSG_SLAB] = "kernel slab page", | |
787 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
cc637b17 XX |
788 | [MF_MSG_HUGE] = "huge page", |
789 | [MF_MSG_FREE_HUGE] = "free huge page", | |
790 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", | |
791 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
792 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
793 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
794 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
795 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
796 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
797 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
798 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
799 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
800 | [MF_MSG_BUDDY] = "free buddy page", | |
6100e34b | 801 | [MF_MSG_DAX] = "dax page", |
5d1fd5dc | 802 | [MF_MSG_UNSPLIT_THP] = "unsplit thp", |
cc637b17 | 803 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
804 | }; |
805 | ||
dc2a1cbf WF |
806 | /* |
807 | * XXX: It is possible that a page is isolated from LRU cache, | |
808 | * and then kept in swap cache or failed to remove from page cache. | |
809 | * The page count will stop it from being freed by unpoison. | |
810 | * Stress tests should be aware of this memory leak problem. | |
811 | */ | |
812 | static int delete_from_lru_cache(struct page *p) | |
813 | { | |
814 | if (!isolate_lru_page(p)) { | |
815 | /* | |
816 | * Clear sensible page flags, so that the buddy system won't | |
817 | * complain when the page is unpoison-and-freed. | |
818 | */ | |
819 | ClearPageActive(p); | |
820 | ClearPageUnevictable(p); | |
18365225 MH |
821 | |
822 | /* | |
823 | * Poisoned page might never drop its ref count to 0 so we have | |
824 | * to uncharge it manually from its memcg. | |
825 | */ | |
bbc6b703 | 826 | mem_cgroup_uncharge(page_folio(p)); |
18365225 | 827 | |
dc2a1cbf WF |
828 | /* |
829 | * drop the page count elevated by isolate_lru_page() | |
830 | */ | |
09cbfeaf | 831 | put_page(p); |
dc2a1cbf WF |
832 | return 0; |
833 | } | |
834 | return -EIO; | |
835 | } | |
836 | ||
78bb9203 NH |
837 | static int truncate_error_page(struct page *p, unsigned long pfn, |
838 | struct address_space *mapping) | |
839 | { | |
840 | int ret = MF_FAILED; | |
841 | ||
842 | if (mapping->a_ops->error_remove_page) { | |
ac5efa78 | 843 | struct folio *folio = page_folio(p); |
78bb9203 NH |
844 | int err = mapping->a_ops->error_remove_page(mapping, p); |
845 | ||
846 | if (err != 0) { | |
96f96763 | 847 | pr_info("%#lx: Failed to punch page: %d\n", pfn, err); |
ac5efa78 VMO |
848 | } else if (folio_has_private(folio) && |
849 | !filemap_release_folio(folio, GFP_NOIO)) { | |
96f96763 | 850 | pr_info("%#lx: failed to release buffers\n", pfn); |
78bb9203 NH |
851 | } else { |
852 | ret = MF_RECOVERED; | |
853 | } | |
854 | } else { | |
855 | /* | |
856 | * If the file system doesn't support it just invalidate | |
857 | * This fails on dirty or anything with private pages | |
858 | */ | |
859 | if (invalidate_inode_page(p)) | |
860 | ret = MF_RECOVERED; | |
861 | else | |
96f96763 | 862 | pr_info("%#lx: Failed to invalidate\n", pfn); |
78bb9203 NH |
863 | } |
864 | ||
865 | return ret; | |
866 | } | |
867 | ||
dd0f230a YS |
868 | struct page_state { |
869 | unsigned long mask; | |
870 | unsigned long res; | |
871 | enum mf_action_page_type type; | |
872 | ||
873 | /* Callback ->action() has to unlock the relevant page inside it. */ | |
874 | int (*action)(struct page_state *ps, struct page *p); | |
875 | }; | |
876 | ||
877 | /* | |
878 | * Return true if page is still referenced by others, otherwise return | |
879 | * false. | |
880 | * | |
881 | * The extra_pins is true when one extra refcount is expected. | |
882 | */ | |
883 | static bool has_extra_refcount(struct page_state *ps, struct page *p, | |
884 | bool extra_pins) | |
885 | { | |
886 | int count = page_count(p) - 1; | |
887 | ||
888 | if (extra_pins) | |
889 | count -= 1; | |
890 | ||
891 | if (count > 0) { | |
96f96763 | 892 | pr_err("%#lx: %s still referenced by %d users\n", |
dd0f230a YS |
893 | page_to_pfn(p), action_page_types[ps->type], count); |
894 | return true; | |
895 | } | |
896 | ||
897 | return false; | |
898 | } | |
899 | ||
6a46079c AK |
900 | /* |
901 | * Error hit kernel page. | |
902 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
903 | * could be more sophisticated. | |
904 | */ | |
dd0f230a | 905 | static int me_kernel(struct page_state *ps, struct page *p) |
6a46079c | 906 | { |
ea6d0630 | 907 | unlock_page(p); |
cc637b17 | 908 | return MF_IGNORED; |
6a46079c AK |
909 | } |
910 | ||
911 | /* | |
912 | * Page in unknown state. Do nothing. | |
913 | */ | |
dd0f230a | 914 | static int me_unknown(struct page_state *ps, struct page *p) |
6a46079c | 915 | { |
96f96763 | 916 | pr_err("%#lx: Unknown page state\n", page_to_pfn(p)); |
ea6d0630 | 917 | unlock_page(p); |
cc637b17 | 918 | return MF_FAILED; |
6a46079c AK |
919 | } |
920 | ||
6a46079c AK |
921 | /* |
922 | * Clean (or cleaned) page cache page. | |
923 | */ | |
dd0f230a | 924 | static int me_pagecache_clean(struct page_state *ps, struct page *p) |
6a46079c | 925 | { |
ea6d0630 | 926 | int ret; |
6a46079c | 927 | struct address_space *mapping; |
a7605426 | 928 | bool extra_pins; |
6a46079c | 929 | |
dc2a1cbf WF |
930 | delete_from_lru_cache(p); |
931 | ||
6a46079c AK |
932 | /* |
933 | * For anonymous pages we're done the only reference left | |
934 | * should be the one m_f() holds. | |
935 | */ | |
ea6d0630 NH |
936 | if (PageAnon(p)) { |
937 | ret = MF_RECOVERED; | |
938 | goto out; | |
939 | } | |
6a46079c AK |
940 | |
941 | /* | |
942 | * Now truncate the page in the page cache. This is really | |
943 | * more like a "temporary hole punch" | |
944 | * Don't do this for block devices when someone else | |
945 | * has a reference, because it could be file system metadata | |
946 | * and that's not safe to truncate. | |
947 | */ | |
948 | mapping = page_mapping(p); | |
949 | if (!mapping) { | |
950 | /* | |
951 | * Page has been teared down in the meanwhile | |
952 | */ | |
ea6d0630 NH |
953 | ret = MF_FAILED; |
954 | goto out; | |
6a46079c AK |
955 | } |
956 | ||
a7605426 YS |
957 | /* |
958 | * The shmem page is kept in page cache instead of truncating | |
959 | * so is expected to have an extra refcount after error-handling. | |
960 | */ | |
961 | extra_pins = shmem_mapping(mapping); | |
962 | ||
6a46079c AK |
963 | /* |
964 | * Truncation is a bit tricky. Enable it per file system for now. | |
965 | * | |
9608703e | 966 | * Open: to take i_rwsem or not for this? Right now we don't. |
6a46079c | 967 | */ |
dd0f230a | 968 | ret = truncate_error_page(p, page_to_pfn(p), mapping); |
a7605426 YS |
969 | if (has_extra_refcount(ps, p, extra_pins)) |
970 | ret = MF_FAILED; | |
971 | ||
ea6d0630 NH |
972 | out: |
973 | unlock_page(p); | |
dd0f230a | 974 | |
ea6d0630 | 975 | return ret; |
6a46079c AK |
976 | } |
977 | ||
978 | /* | |
549543df | 979 | * Dirty pagecache page |
6a46079c AK |
980 | * Issues: when the error hit a hole page the error is not properly |
981 | * propagated. | |
982 | */ | |
dd0f230a | 983 | static int me_pagecache_dirty(struct page_state *ps, struct page *p) |
6a46079c AK |
984 | { |
985 | struct address_space *mapping = page_mapping(p); | |
986 | ||
987 | SetPageError(p); | |
988 | /* TBD: print more information about the file. */ | |
989 | if (mapping) { | |
990 | /* | |
991 | * IO error will be reported by write(), fsync(), etc. | |
992 | * who check the mapping. | |
993 | * This way the application knows that something went | |
994 | * wrong with its dirty file data. | |
995 | * | |
996 | * There's one open issue: | |
997 | * | |
998 | * The EIO will be only reported on the next IO | |
999 | * operation and then cleared through the IO map. | |
1000 | * Normally Linux has two mechanisms to pass IO error | |
1001 | * first through the AS_EIO flag in the address space | |
1002 | * and then through the PageError flag in the page. | |
1003 | * Since we drop pages on memory failure handling the | |
1004 | * only mechanism open to use is through AS_AIO. | |
1005 | * | |
1006 | * This has the disadvantage that it gets cleared on | |
1007 | * the first operation that returns an error, while | |
1008 | * the PageError bit is more sticky and only cleared | |
1009 | * when the page is reread or dropped. If an | |
1010 | * application assumes it will always get error on | |
1011 | * fsync, but does other operations on the fd before | |
25985edc | 1012 | * and the page is dropped between then the error |
6a46079c AK |
1013 | * will not be properly reported. |
1014 | * | |
1015 | * This can already happen even without hwpoisoned | |
1016 | * pages: first on metadata IO errors (which only | |
1017 | * report through AS_EIO) or when the page is dropped | |
1018 | * at the wrong time. | |
1019 | * | |
1020 | * So right now we assume that the application DTRT on | |
1021 | * the first EIO, but we're not worse than other parts | |
1022 | * of the kernel. | |
1023 | */ | |
af21bfaf | 1024 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
1025 | } |
1026 | ||
dd0f230a | 1027 | return me_pagecache_clean(ps, p); |
6a46079c AK |
1028 | } |
1029 | ||
1030 | /* | |
1031 | * Clean and dirty swap cache. | |
1032 | * | |
1033 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
1034 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
1035 | * referenced concurrently by 2 types of PTEs: | |
1036 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
1037 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
1038 | * and then | |
1039 | * - clear dirty bit to prevent IO | |
1040 | * - remove from LRU | |
1041 | * - but keep in the swap cache, so that when we return to it on | |
1042 | * a later page fault, we know the application is accessing | |
1043 | * corrupted data and shall be killed (we installed simple | |
1044 | * interception code in do_swap_page to catch it). | |
1045 | * | |
1046 | * Clean swap cache pages can be directly isolated. A later page fault will | |
1047 | * bring in the known good data from disk. | |
1048 | */ | |
dd0f230a | 1049 | static int me_swapcache_dirty(struct page_state *ps, struct page *p) |
6a46079c | 1050 | { |
ea6d0630 | 1051 | int ret; |
dd0f230a | 1052 | bool extra_pins = false; |
ea6d0630 | 1053 | |
6a46079c AK |
1054 | ClearPageDirty(p); |
1055 | /* Trigger EIO in shmem: */ | |
1056 | ClearPageUptodate(p); | |
1057 | ||
ea6d0630 NH |
1058 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED; |
1059 | unlock_page(p); | |
dd0f230a YS |
1060 | |
1061 | if (ret == MF_DELAYED) | |
1062 | extra_pins = true; | |
1063 | ||
1064 | if (has_extra_refcount(ps, p, extra_pins)) | |
1065 | ret = MF_FAILED; | |
1066 | ||
ea6d0630 | 1067 | return ret; |
6a46079c AK |
1068 | } |
1069 | ||
dd0f230a | 1070 | static int me_swapcache_clean(struct page_state *ps, struct page *p) |
6a46079c | 1071 | { |
75fa68a5 | 1072 | struct folio *folio = page_folio(p); |
ea6d0630 NH |
1073 | int ret; |
1074 | ||
75fa68a5 | 1075 | delete_from_swap_cache(folio); |
e43c3afb | 1076 | |
ea6d0630 | 1077 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED; |
75fa68a5 | 1078 | folio_unlock(folio); |
dd0f230a YS |
1079 | |
1080 | if (has_extra_refcount(ps, p, false)) | |
1081 | ret = MF_FAILED; | |
1082 | ||
ea6d0630 | 1083 | return ret; |
6a46079c AK |
1084 | } |
1085 | ||
1086 | /* | |
1087 | * Huge pages. Needs work. | |
1088 | * Issues: | |
93f70f90 NH |
1089 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
1090 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c | 1091 | */ |
dd0f230a | 1092 | static int me_huge_page(struct page_state *ps, struct page *p) |
6a46079c | 1093 | { |
a8b2c2ce | 1094 | int res; |
93f70f90 | 1095 | struct page *hpage = compound_head(p); |
78bb9203 | 1096 | struct address_space *mapping; |
8625147c | 1097 | bool extra_pins = false; |
2491ffee NH |
1098 | |
1099 | if (!PageHuge(hpage)) | |
1100 | return MF_DELAYED; | |
1101 | ||
78bb9203 NH |
1102 | mapping = page_mapping(hpage); |
1103 | if (mapping) { | |
dd0f230a | 1104 | res = truncate_error_page(hpage, page_to_pfn(p), mapping); |
8625147c JH |
1105 | /* The page is kept in page cache. */ |
1106 | extra_pins = true; | |
ea6d0630 | 1107 | unlock_page(hpage); |
78bb9203 NH |
1108 | } else { |
1109 | unlock_page(hpage); | |
1110 | /* | |
ef526b17 ML |
1111 | * migration entry prevents later access on error hugepage, |
1112 | * so we can free and dissolve it into buddy to save healthy | |
1113 | * subpages. | |
78bb9203 | 1114 | */ |
ef526b17 | 1115 | put_page(hpage); |
ceaf8fbe | 1116 | if (__page_handle_poison(p) >= 0) { |
a8b2c2ce OS |
1117 | page_ref_inc(p); |
1118 | res = MF_RECOVERED; | |
ceaf8fbe NH |
1119 | } else { |
1120 | res = MF_FAILED; | |
a8b2c2ce | 1121 | } |
93f70f90 | 1122 | } |
78bb9203 | 1123 | |
8625147c | 1124 | if (has_extra_refcount(ps, p, extra_pins)) |
dd0f230a YS |
1125 | res = MF_FAILED; |
1126 | ||
78bb9203 | 1127 | return res; |
6a46079c AK |
1128 | } |
1129 | ||
1130 | /* | |
1131 | * Various page states we can handle. | |
1132 | * | |
1133 | * A page state is defined by its current page->flags bits. | |
1134 | * The table matches them in order and calls the right handler. | |
1135 | * | |
1136 | * This is quite tricky because we can access page at any time | |
25985edc | 1137 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
1138 | * |
1139 | * This is not complete. More states could be added. | |
1140 | * For any missing state don't attempt recovery. | |
1141 | */ | |
1142 | ||
1143 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 1144 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
1145 | #define unevict (1UL << PG_unevictable) |
1146 | #define mlock (1UL << PG_mlocked) | |
6a46079c | 1147 | #define lru (1UL << PG_lru) |
6a46079c | 1148 | #define head (1UL << PG_head) |
6a46079c | 1149 | #define slab (1UL << PG_slab) |
6a46079c AK |
1150 | #define reserved (1UL << PG_reserved) |
1151 | ||
dd0f230a | 1152 | static struct page_state error_states[] = { |
cc637b17 | 1153 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
1154 | /* |
1155 | * free pages are specially detected outside this table: | |
1156 | * PG_buddy pages only make a small fraction of all free pages. | |
1157 | */ | |
6a46079c AK |
1158 | |
1159 | /* | |
1160 | * Could in theory check if slab page is free or if we can drop | |
1161 | * currently unused objects without touching them. But just | |
1162 | * treat it as standard kernel for now. | |
1163 | */ | |
cc637b17 | 1164 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 1165 | |
cc637b17 | 1166 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 1167 | |
cc637b17 XX |
1168 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
1169 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 1170 | |
cc637b17 XX |
1171 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
1172 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 1173 | |
cc637b17 XX |
1174 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
1175 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 1176 | |
cc637b17 XX |
1177 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
1178 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
1179 | |
1180 | /* | |
1181 | * Catchall entry: must be at end. | |
1182 | */ | |
cc637b17 | 1183 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
1184 | }; |
1185 | ||
2326c467 AK |
1186 | #undef dirty |
1187 | #undef sc | |
1188 | #undef unevict | |
1189 | #undef mlock | |
2326c467 | 1190 | #undef lru |
2326c467 | 1191 | #undef head |
2326c467 AK |
1192 | #undef slab |
1193 | #undef reserved | |
1194 | ||
ff604cf6 NH |
1195 | /* |
1196 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
1197 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
1198 | */ | |
b66d00df KW |
1199 | static int action_result(unsigned long pfn, enum mf_action_page_type type, |
1200 | enum mf_result result) | |
6a46079c | 1201 | { |
97f0b134 XX |
1202 | trace_memory_failure_event(pfn, type, result); |
1203 | ||
a46c9304 | 1204 | num_poisoned_pages_inc(pfn); |
96f96763 | 1205 | pr_err("%#lx: recovery action for %s: %s\n", |
64d37a2b | 1206 | pfn, action_page_types[type], action_name[result]); |
b66d00df KW |
1207 | |
1208 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; | |
6a46079c AK |
1209 | } |
1210 | ||
1211 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 1212 | unsigned long pfn) |
6a46079c AK |
1213 | { |
1214 | int result; | |
1215 | ||
ea6d0630 | 1216 | /* page p should be unlocked after returning from ps->action(). */ |
dd0f230a | 1217 | result = ps->action(ps, p); |
7456b040 | 1218 | |
6a46079c AK |
1219 | /* Could do more checks here if page looks ok */ |
1220 | /* | |
1221 | * Could adjust zone counters here to correct for the missing page. | |
1222 | */ | |
1223 | ||
b66d00df | 1224 | return action_result(pfn, ps->type, result); |
6a46079c AK |
1225 | } |
1226 | ||
bf181c58 NH |
1227 | static inline bool PageHWPoisonTakenOff(struct page *page) |
1228 | { | |
1229 | return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON; | |
1230 | } | |
1231 | ||
1232 | void SetPageHWPoisonTakenOff(struct page *page) | |
1233 | { | |
1234 | set_page_private(page, MAGIC_HWPOISON); | |
1235 | } | |
1236 | ||
1237 | void ClearPageHWPoisonTakenOff(struct page *page) | |
1238 | { | |
1239 | if (PageHWPoison(page)) | |
1240 | set_page_private(page, 0); | |
1241 | } | |
1242 | ||
25182f05 NH |
1243 | /* |
1244 | * Return true if a page type of a given page is supported by hwpoison | |
1245 | * mechanism (while handling could fail), otherwise false. This function | |
1246 | * does not return true for hugetlb or device memory pages, so it's assumed | |
1247 | * to be called only in the context where we never have such pages. | |
1248 | */ | |
bf6445bc | 1249 | static inline bool HWPoisonHandlable(struct page *page, unsigned long flags) |
25182f05 | 1250 | { |
3f871370 | 1251 | /* Soft offline could migrate non-LRU movable pages */ |
bf6445bc | 1252 | if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page)) |
3f871370 | 1253 | return true; |
bf6445bc | 1254 | |
3f871370 | 1255 | return PageLRU(page) || is_free_buddy_page(page); |
25182f05 NH |
1256 | } |
1257 | ||
bf6445bc | 1258 | static int __get_hwpoison_page(struct page *page, unsigned long flags) |
ead07f6a NH |
1259 | { |
1260 | struct page *head = compound_head(page); | |
25182f05 NH |
1261 | int ret = 0; |
1262 | bool hugetlb = false; | |
1263 | ||
e591ef7d | 1264 | ret = get_hwpoison_huge_page(head, &hugetlb, false); |
25182f05 NH |
1265 | if (hugetlb) |
1266 | return ret; | |
1267 | ||
1268 | /* | |
9cf28191 ML |
1269 | * This check prevents from calling get_page_unless_zero() for any |
1270 | * unsupported type of page in order to reduce the risk of unexpected | |
1271 | * races caused by taking a page refcount. | |
25182f05 | 1272 | */ |
bf6445bc | 1273 | if (!HWPoisonHandlable(head, flags)) |
fcc00621 | 1274 | return -EBUSY; |
ead07f6a | 1275 | |
c2e7e00b KK |
1276 | if (get_page_unless_zero(head)) { |
1277 | if (head == compound_head(page)) | |
1278 | return 1; | |
1279 | ||
96f96763 | 1280 | pr_info("%#lx cannot catch tail\n", page_to_pfn(page)); |
c2e7e00b KK |
1281 | put_page(head); |
1282 | } | |
1283 | ||
1284 | return 0; | |
ead07f6a | 1285 | } |
ead07f6a | 1286 | |
2f714160 | 1287 | static int get_any_page(struct page *p, unsigned long flags) |
17e395b6 | 1288 | { |
2f714160 OS |
1289 | int ret = 0, pass = 0; |
1290 | bool count_increased = false; | |
17e395b6 | 1291 | |
2f714160 OS |
1292 | if (flags & MF_COUNT_INCREASED) |
1293 | count_increased = true; | |
1294 | ||
1295 | try_again: | |
0ed950d1 | 1296 | if (!count_increased) { |
bf6445bc | 1297 | ret = __get_hwpoison_page(p, flags); |
0ed950d1 NH |
1298 | if (!ret) { |
1299 | if (page_count(p)) { | |
1300 | /* We raced with an allocation, retry. */ | |
1301 | if (pass++ < 3) | |
1302 | goto try_again; | |
1303 | ret = -EBUSY; | |
1304 | } else if (!PageHuge(p) && !is_free_buddy_page(p)) { | |
1305 | /* We raced with put_page, retry. */ | |
1306 | if (pass++ < 3) | |
1307 | goto try_again; | |
1308 | ret = -EIO; | |
1309 | } | |
1310 | goto out; | |
1311 | } else if (ret == -EBUSY) { | |
fcc00621 NH |
1312 | /* |
1313 | * We raced with (possibly temporary) unhandlable | |
1314 | * page, retry. | |
1315 | */ | |
1316 | if (pass++ < 3) { | |
d0505e9f | 1317 | shake_page(p); |
2f714160 | 1318 | goto try_again; |
fcc00621 NH |
1319 | } |
1320 | ret = -EIO; | |
0ed950d1 | 1321 | goto out; |
2f714160 | 1322 | } |
0ed950d1 NH |
1323 | } |
1324 | ||
bf6445bc | 1325 | if (PageHuge(p) || HWPoisonHandlable(p, flags)) { |
0ed950d1 | 1326 | ret = 1; |
2f714160 | 1327 | } else { |
0ed950d1 NH |
1328 | /* |
1329 | * A page we cannot handle. Check whether we can turn | |
1330 | * it into something we can handle. | |
1331 | */ | |
1332 | if (pass++ < 3) { | |
2f714160 | 1333 | put_page(p); |
d0505e9f | 1334 | shake_page(p); |
0ed950d1 NH |
1335 | count_increased = false; |
1336 | goto try_again; | |
2f714160 | 1337 | } |
0ed950d1 NH |
1338 | put_page(p); |
1339 | ret = -EIO; | |
17e395b6 | 1340 | } |
0ed950d1 | 1341 | out: |
941ca063 | 1342 | if (ret == -EIO) |
96f96763 | 1343 | pr_err("%#lx: unhandlable page.\n", page_to_pfn(p)); |
941ca063 | 1344 | |
17e395b6 OS |
1345 | return ret; |
1346 | } | |
1347 | ||
bf181c58 NH |
1348 | static int __get_unpoison_page(struct page *page) |
1349 | { | |
1350 | struct page *head = compound_head(page); | |
1351 | int ret = 0; | |
1352 | bool hugetlb = false; | |
1353 | ||
e591ef7d | 1354 | ret = get_hwpoison_huge_page(head, &hugetlb, true); |
bf181c58 NH |
1355 | if (hugetlb) |
1356 | return ret; | |
1357 | ||
1358 | /* | |
1359 | * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison, | |
1360 | * but also isolated from buddy freelist, so need to identify the | |
1361 | * state and have to cancel both operations to unpoison. | |
1362 | */ | |
1363 | if (PageHWPoisonTakenOff(page)) | |
1364 | return -EHWPOISON; | |
1365 | ||
1366 | return get_page_unless_zero(page) ? 1 : 0; | |
1367 | } | |
1368 | ||
0ed950d1 NH |
1369 | /** |
1370 | * get_hwpoison_page() - Get refcount for memory error handling | |
1371 | * @p: Raw error page (hit by memory error) | |
1372 | * @flags: Flags controlling behavior of error handling | |
1373 | * | |
1374 | * get_hwpoison_page() takes a page refcount of an error page to handle memory | |
1375 | * error on it, after checking that the error page is in a well-defined state | |
0b8f0d87 | 1376 | * (defined as a page-type we can successfully handle the memory error on it, |
0ed950d1 NH |
1377 | * such as LRU page and hugetlb page). |
1378 | * | |
1379 | * Memory error handling could be triggered at any time on any type of page, | |
1380 | * so it's prone to race with typical memory management lifecycle (like | |
1381 | * allocation and free). So to avoid such races, get_hwpoison_page() takes | |
1382 | * extra care for the error page's state (as done in __get_hwpoison_page()), | |
1383 | * and has some retry logic in get_any_page(). | |
1384 | * | |
bf181c58 NH |
1385 | * When called from unpoison_memory(), the caller should already ensure that |
1386 | * the given page has PG_hwpoison. So it's never reused for other page | |
1387 | * allocations, and __get_unpoison_page() never races with them. | |
1388 | * | |
0ed950d1 NH |
1389 | * Return: 0 on failure, |
1390 | * 1 on success for in-use pages in a well-defined state, | |
1391 | * -EIO for pages on which we can not handle memory errors, | |
1392 | * -EBUSY when get_hwpoison_page() has raced with page lifecycle | |
bf181c58 NH |
1393 | * operations like allocation and free, |
1394 | * -EHWPOISON when the page is hwpoisoned and taken off from buddy. | |
0ed950d1 NH |
1395 | */ |
1396 | static int get_hwpoison_page(struct page *p, unsigned long flags) | |
2f714160 OS |
1397 | { |
1398 | int ret; | |
1399 | ||
1400 | zone_pcp_disable(page_zone(p)); | |
bf181c58 NH |
1401 | if (flags & MF_UNPOISON) |
1402 | ret = __get_unpoison_page(p); | |
1403 | else | |
1404 | ret = get_any_page(p, flags); | |
2f714160 OS |
1405 | zone_pcp_enable(page_zone(p)); |
1406 | ||
1407 | return ret; | |
1408 | } | |
1409 | ||
6a46079c AK |
1410 | /* |
1411 | * Do all that is necessary to remove user space mappings. Unmap | |
1412 | * the pages and send SIGBUS to the processes if the data was dirty. | |
1413 | */ | |
666e5a40 | 1414 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
ed8c2f49 | 1415 | int flags, struct page *hpage) |
6a46079c | 1416 | { |
869f7ee6 | 1417 | struct folio *folio = page_folio(hpage); |
36af6737 | 1418 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC; |
6a46079c AK |
1419 | struct address_space *mapping; |
1420 | LIST_HEAD(tokill); | |
1fb08ac6 | 1421 | bool unmap_success; |
0792a4a6 | 1422 | int forcekill; |
286c469a | 1423 | bool mlocked = PageMlocked(hpage); |
6a46079c | 1424 | |
93a9eb39 NH |
1425 | /* |
1426 | * Here we are interested only in user-mapped pages, so skip any | |
1427 | * other types of pages. | |
1428 | */ | |
b680dae9 | 1429 | if (PageReserved(p) || PageSlab(p) || PageTable(p)) |
666e5a40 | 1430 | return true; |
93a9eb39 | 1431 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 1432 | return true; |
6a46079c | 1433 | |
6a46079c AK |
1434 | /* |
1435 | * This check implies we don't kill processes if their pages | |
1436 | * are in the swap cache early. Those are always late kills. | |
1437 | */ | |
7af446a8 | 1438 | if (!page_mapped(hpage)) |
666e5a40 | 1439 | return true; |
1668bfd5 | 1440 | |
52089b14 | 1441 | if (PageKsm(p)) { |
96f96763 | 1442 | pr_err("%#lx: can't handle KSM pages.\n", pfn); |
666e5a40 | 1443 | return false; |
52089b14 | 1444 | } |
6a46079c AK |
1445 | |
1446 | if (PageSwapCache(p)) { | |
96f96763 | 1447 | pr_err("%#lx: keeping poisoned page in swap cache\n", pfn); |
6a46079c AK |
1448 | ttu |= TTU_IGNORE_HWPOISON; |
1449 | } | |
1450 | ||
1451 | /* | |
1452 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1453 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1454 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1455 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1456 | */ |
7af446a8 | 1457 | mapping = page_mapping(hpage); |
6751ed65 | 1458 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
f56753ac | 1459 | mapping_can_writeback(mapping)) { |
7af446a8 NH |
1460 | if (page_mkclean(hpage)) { |
1461 | SetPageDirty(hpage); | |
6a46079c | 1462 | } else { |
6a46079c | 1463 | ttu |= TTU_IGNORE_HWPOISON; |
96f96763 | 1464 | pr_info("%#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1465 | pfn); |
1466 | } | |
1467 | } | |
1468 | ||
1469 | /* | |
1470 | * First collect all the processes that have the page | |
1471 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1472 | * because ttu takes the rmap data structures down. | |
6a46079c | 1473 | */ |
0792a4a6 | 1474 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1475 | |
357670f7 ML |
1476 | if (PageHuge(hpage) && !PageAnon(hpage)) { |
1477 | /* | |
1478 | * For hugetlb pages in shared mappings, try_to_unmap | |
1479 | * could potentially call huge_pmd_unshare. Because of | |
1480 | * this, take semaphore in write mode here and set | |
1481 | * TTU_RMAP_LOCKED to indicate we have taken the lock | |
1482 | * at this higher level. | |
1483 | */ | |
1484 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1485 | if (mapping) { | |
9030fb0b | 1486 | try_to_unmap(folio, ttu|TTU_RMAP_LOCKED); |
357670f7 ML |
1487 | i_mmap_unlock_write(mapping); |
1488 | } else | |
96f96763 | 1489 | pr_info("%#lx: could not lock mapping for mapped huge page\n", pfn); |
c0d0381a | 1490 | } else { |
9030fb0b | 1491 | try_to_unmap(folio, ttu); |
c0d0381a | 1492 | } |
1fb08ac6 YS |
1493 | |
1494 | unmap_success = !page_mapped(hpage); | |
666e5a40 | 1495 | if (!unmap_success) |
96f96763 | 1496 | pr_err("%#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1497 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1498 | |
286c469a NH |
1499 | /* |
1500 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1501 | * shake_page() again to ensure that it's flushed. | |
1502 | */ | |
1503 | if (mlocked) | |
d0505e9f | 1504 | shake_page(hpage); |
286c469a | 1505 | |
6a46079c AK |
1506 | /* |
1507 | * Now that the dirty bit has been propagated to the | |
1508 | * struct page and all unmaps done we can decide if | |
1509 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1510 | * was dirty or the process is not restartable, |
1511 | * otherwise the tokill list is merely | |
6a46079c AK |
1512 | * freed. When there was a problem unmapping earlier |
1513 | * use a more force-full uncatchable kill to prevent | |
1514 | * any accesses to the poisoned memory. | |
1515 | */ | |
0792a4a6 ML |
1516 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL) || |
1517 | !unmap_success; | |
ae1139ec | 1518 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1519 | |
666e5a40 | 1520 | return unmap_success; |
6a46079c AK |
1521 | } |
1522 | ||
0348d2eb NH |
1523 | static int identify_page_state(unsigned long pfn, struct page *p, |
1524 | unsigned long page_flags) | |
761ad8d7 NH |
1525 | { |
1526 | struct page_state *ps; | |
0348d2eb NH |
1527 | |
1528 | /* | |
1529 | * The first check uses the current page flags which may not have any | |
1530 | * relevant information. The second check with the saved page flags is | |
1531 | * carried out only if the first check can't determine the page status. | |
1532 | */ | |
1533 | for (ps = error_states;; ps++) | |
1534 | if ((p->flags & ps->mask) == ps->res) | |
1535 | break; | |
1536 | ||
1537 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1538 | ||
1539 | if (!ps->mask) | |
1540 | for (ps = error_states;; ps++) | |
1541 | if ((page_flags & ps->mask) == ps->res) | |
1542 | break; | |
1543 | return page_action(ps, p, pfn); | |
1544 | } | |
1545 | ||
2ace36f0 | 1546 | static int try_to_split_thp_page(struct page *page) |
694bf0b0 | 1547 | { |
2ace36f0 KW |
1548 | int ret; |
1549 | ||
694bf0b0 | 1550 | lock_page(page); |
2ace36f0 KW |
1551 | ret = split_huge_page(page); |
1552 | unlock_page(page); | |
694bf0b0 | 1553 | |
2ace36f0 | 1554 | if (unlikely(ret)) |
694bf0b0 | 1555 | put_page(page); |
694bf0b0 | 1556 | |
2ace36f0 | 1557 | return ret; |
694bf0b0 OS |
1558 | } |
1559 | ||
00cc790e SR |
1560 | static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn, |
1561 | struct address_space *mapping, pgoff_t index, int flags) | |
1562 | { | |
1563 | struct to_kill *tk; | |
1564 | unsigned long size = 0; | |
1565 | ||
1566 | list_for_each_entry(tk, to_kill, nd) | |
1567 | if (tk->size_shift) | |
1568 | size = max(size, 1UL << tk->size_shift); | |
1569 | ||
1570 | if (size) { | |
1571 | /* | |
1572 | * Unmap the largest mapping to avoid breaking up device-dax | |
1573 | * mappings which are constant size. The actual size of the | |
1574 | * mapping being torn down is communicated in siginfo, see | |
1575 | * kill_proc() | |
1576 | */ | |
1577 | loff_t start = (index << PAGE_SHIFT) & ~(size - 1); | |
1578 | ||
1579 | unmap_mapping_range(mapping, start, size, 0); | |
1580 | } | |
1581 | ||
1582 | kill_procs(to_kill, flags & MF_MUST_KILL, false, pfn, flags); | |
1583 | } | |
1584 | ||
1585 | static int mf_generic_kill_procs(unsigned long long pfn, int flags, | |
1586 | struct dev_pagemap *pgmap) | |
1587 | { | |
1588 | struct page *page = pfn_to_page(pfn); | |
1589 | LIST_HEAD(to_kill); | |
1590 | dax_entry_t cookie; | |
1591 | int rc = 0; | |
1592 | ||
1593 | /* | |
1594 | * Pages instantiated by device-dax (not filesystem-dax) | |
1595 | * may be compound pages. | |
1596 | */ | |
1597 | page = compound_head(page); | |
1598 | ||
1599 | /* | |
1600 | * Prevent the inode from being freed while we are interrogating | |
1601 | * the address_space, typically this would be handled by | |
1602 | * lock_page(), but dax pages do not use the page lock. This | |
1603 | * also prevents changes to the mapping of this pfn until | |
1604 | * poison signaling is complete. | |
1605 | */ | |
1606 | cookie = dax_lock_page(page); | |
1607 | if (!cookie) | |
1608 | return -EBUSY; | |
1609 | ||
1610 | if (hwpoison_filter(page)) { | |
1611 | rc = -EOPNOTSUPP; | |
1612 | goto unlock; | |
1613 | } | |
1614 | ||
1615 | switch (pgmap->type) { | |
1616 | case MEMORY_DEVICE_PRIVATE: | |
1617 | case MEMORY_DEVICE_COHERENT: | |
1618 | /* | |
1619 | * TODO: Handle device pages which may need coordination | |
1620 | * with device-side memory. | |
1621 | */ | |
1622 | rc = -ENXIO; | |
1623 | goto unlock; | |
1624 | default: | |
1625 | break; | |
1626 | } | |
1627 | ||
1628 | /* | |
1629 | * Use this flag as an indication that the dax page has been | |
1630 | * remapped UC to prevent speculative consumption of poison. | |
1631 | */ | |
1632 | SetPageHWPoison(page); | |
1633 | ||
1634 | /* | |
1635 | * Unlike System-RAM there is no possibility to swap in a | |
1636 | * different physical page at a given virtual address, so all | |
1637 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1638 | * SIGBUS (i.e. MF_MUST_KILL) | |
1639 | */ | |
1640 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1641 | collect_procs(page, &to_kill, true); | |
1642 | ||
1643 | unmap_and_kill(&to_kill, pfn, page->mapping, page->index, flags); | |
1644 | unlock: | |
1645 | dax_unlock_page(page, cookie); | |
1646 | return rc; | |
1647 | } | |
1648 | ||
c36e2024 SR |
1649 | #ifdef CONFIG_FS_DAX |
1650 | /** | |
1651 | * mf_dax_kill_procs - Collect and kill processes who are using this file range | |
1652 | * @mapping: address_space of the file in use | |
1653 | * @index: start pgoff of the range within the file | |
1654 | * @count: length of the range, in unit of PAGE_SIZE | |
1655 | * @mf_flags: memory failure flags | |
1656 | */ | |
1657 | int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index, | |
1658 | unsigned long count, int mf_flags) | |
1659 | { | |
1660 | LIST_HEAD(to_kill); | |
1661 | dax_entry_t cookie; | |
1662 | struct page *page; | |
1663 | size_t end = index + count; | |
1664 | ||
1665 | mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1666 | ||
1667 | for (; index < end; index++) { | |
1668 | page = NULL; | |
1669 | cookie = dax_lock_mapping_entry(mapping, index, &page); | |
1670 | if (!cookie) | |
1671 | return -EBUSY; | |
1672 | if (!page) | |
1673 | goto unlock; | |
1674 | ||
1675 | SetPageHWPoison(page); | |
1676 | ||
1677 | collect_procs_fsdax(page, mapping, index, &to_kill); | |
1678 | unmap_and_kill(&to_kill, page_to_pfn(page), mapping, | |
1679 | index, mf_flags); | |
1680 | unlock: | |
1681 | dax_unlock_mapping_entry(mapping, index, cookie); | |
1682 | } | |
1683 | return 0; | |
1684 | } | |
1685 | EXPORT_SYMBOL_GPL(mf_dax_kill_procs); | |
1686 | #endif /* CONFIG_FS_DAX */ | |
1687 | ||
161df60e NH |
1688 | #ifdef CONFIG_HUGETLB_PAGE |
1689 | /* | |
1690 | * Struct raw_hwp_page represents information about "raw error page", | |
dad6a5eb | 1691 | * constructing singly linked list from ->_hugetlb_hwpoison field of folio. |
161df60e NH |
1692 | */ |
1693 | struct raw_hwp_page { | |
1694 | struct llist_node node; | |
1695 | struct page *page; | |
1696 | }; | |
1697 | ||
1698 | static inline struct llist_head *raw_hwp_list_head(struct page *hpage) | |
1699 | { | |
dad6a5eb | 1700 | return (struct llist_head *)&page_folio(hpage)->_hugetlb_hwpoison; |
161df60e NH |
1701 | } |
1702 | ||
ac5fcde0 | 1703 | static unsigned long __free_raw_hwp_pages(struct page *hpage, bool move_flag) |
161df60e NH |
1704 | { |
1705 | struct llist_head *head; | |
1706 | struct llist_node *t, *tnode; | |
ac5fcde0 | 1707 | unsigned long count = 0; |
161df60e NH |
1708 | |
1709 | head = raw_hwp_list_head(hpage); | |
1710 | llist_for_each_safe(tnode, t, head->first) { | |
1711 | struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node); | |
1712 | ||
ac5fcde0 NH |
1713 | if (move_flag) |
1714 | SetPageHWPoison(p->page); | |
5033091d NH |
1715 | else |
1716 | num_poisoned_pages_sub(page_to_pfn(p->page), 1); | |
161df60e | 1717 | kfree(p); |
ac5fcde0 | 1718 | count++; |
161df60e NH |
1719 | } |
1720 | llist_del_all(head); | |
ac5fcde0 | 1721 | return count; |
161df60e NH |
1722 | } |
1723 | ||
1724 | static int hugetlb_set_page_hwpoison(struct page *hpage, struct page *page) | |
1725 | { | |
1726 | struct llist_head *head; | |
1727 | struct raw_hwp_page *raw_hwp; | |
1728 | struct llist_node *t, *tnode; | |
1729 | int ret = TestSetPageHWPoison(hpage) ? -EHWPOISON : 0; | |
1730 | ||
1731 | /* | |
1732 | * Once the hwpoison hugepage has lost reliable raw error info, | |
1733 | * there is little meaning to keep additional error info precisely, | |
1734 | * so skip to add additional raw error info. | |
1735 | */ | |
1736 | if (HPageRawHwpUnreliable(hpage)) | |
1737 | return -EHWPOISON; | |
1738 | head = raw_hwp_list_head(hpage); | |
1739 | llist_for_each_safe(tnode, t, head->first) { | |
1740 | struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node); | |
1741 | ||
1742 | if (p->page == page) | |
1743 | return -EHWPOISON; | |
1744 | } | |
1745 | ||
1746 | raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC); | |
1747 | if (raw_hwp) { | |
1748 | raw_hwp->page = page; | |
1749 | llist_add(&raw_hwp->node, head); | |
1750 | /* the first error event will be counted in action_result(). */ | |
1751 | if (ret) | |
a46c9304 | 1752 | num_poisoned_pages_inc(page_to_pfn(page)); |
161df60e NH |
1753 | } else { |
1754 | /* | |
1755 | * Failed to save raw error info. We no longer trace all | |
1756 | * hwpoisoned subpages, and we need refuse to free/dissolve | |
1757 | * this hwpoisoned hugepage. | |
1758 | */ | |
1759 | SetHPageRawHwpUnreliable(hpage); | |
1760 | /* | |
1761 | * Once HPageRawHwpUnreliable is set, raw_hwp_page is not | |
1762 | * used any more, so free it. | |
1763 | */ | |
ac5fcde0 | 1764 | __free_raw_hwp_pages(hpage, false); |
161df60e NH |
1765 | } |
1766 | return ret; | |
1767 | } | |
1768 | ||
ac5fcde0 NH |
1769 | static unsigned long free_raw_hwp_pages(struct page *hpage, bool move_flag) |
1770 | { | |
1771 | /* | |
1772 | * HPageVmemmapOptimized hugepages can't be freed because struct | |
1773 | * pages for tail pages are required but they don't exist. | |
1774 | */ | |
1775 | if (move_flag && HPageVmemmapOptimized(hpage)) | |
1776 | return 0; | |
1777 | ||
1778 | /* | |
1779 | * HPageRawHwpUnreliable hugepages shouldn't be unpoisoned by | |
1780 | * definition. | |
1781 | */ | |
1782 | if (HPageRawHwpUnreliable(hpage)) | |
1783 | return 0; | |
1784 | ||
1785 | return __free_raw_hwp_pages(hpage, move_flag); | |
1786 | } | |
1787 | ||
161df60e NH |
1788 | void hugetlb_clear_page_hwpoison(struct page *hpage) |
1789 | { | |
1790 | if (HPageRawHwpUnreliable(hpage)) | |
1791 | return; | |
1792 | ClearPageHWPoison(hpage); | |
ac5fcde0 | 1793 | free_raw_hwp_pages(hpage, true); |
161df60e NH |
1794 | } |
1795 | ||
405ce051 NH |
1796 | /* |
1797 | * Called from hugetlb code with hugetlb_lock held. | |
1798 | * | |
1799 | * Return values: | |
1800 | * 0 - free hugepage | |
1801 | * 1 - in-use hugepage | |
1802 | * 2 - not a hugepage | |
1803 | * -EBUSY - the hugepage is busy (try to retry) | |
1804 | * -EHWPOISON - the hugepage is already hwpoisoned | |
1805 | */ | |
e591ef7d NH |
1806 | int __get_huge_page_for_hwpoison(unsigned long pfn, int flags, |
1807 | bool *migratable_cleared) | |
405ce051 NH |
1808 | { |
1809 | struct page *page = pfn_to_page(pfn); | |
1810 | struct page *head = compound_head(page); | |
1811 | int ret = 2; /* fallback to normal page handling */ | |
1812 | bool count_increased = false; | |
1813 | ||
1814 | if (!PageHeadHuge(head)) | |
1815 | goto out; | |
1816 | ||
1817 | if (flags & MF_COUNT_INCREASED) { | |
1818 | ret = 1; | |
1819 | count_increased = true; | |
b283d983 NH |
1820 | } else if (HPageFreed(head)) { |
1821 | ret = 0; | |
1822 | } else if (HPageMigratable(head)) { | |
405ce051 NH |
1823 | ret = get_page_unless_zero(head); |
1824 | if (ret) | |
1825 | count_increased = true; | |
1826 | } else { | |
1827 | ret = -EBUSY; | |
38f6d293 NH |
1828 | if (!(flags & MF_NO_RETRY)) |
1829 | goto out; | |
405ce051 NH |
1830 | } |
1831 | ||
161df60e | 1832 | if (hugetlb_set_page_hwpoison(head, page)) { |
405ce051 NH |
1833 | ret = -EHWPOISON; |
1834 | goto out; | |
1835 | } | |
1836 | ||
e591ef7d NH |
1837 | /* |
1838 | * Clearing HPageMigratable for hwpoisoned hugepages to prevent them | |
1839 | * from being migrated by memory hotremove. | |
1840 | */ | |
1841 | if (count_increased && HPageMigratable(head)) { | |
1842 | ClearHPageMigratable(head); | |
1843 | *migratable_cleared = true; | |
1844 | } | |
1845 | ||
405ce051 NH |
1846 | return ret; |
1847 | out: | |
1848 | if (count_increased) | |
1849 | put_page(head); | |
1850 | return ret; | |
1851 | } | |
1852 | ||
405ce051 NH |
1853 | /* |
1854 | * Taking refcount of hugetlb pages needs extra care about race conditions | |
1855 | * with basic operations like hugepage allocation/free/demotion. | |
1856 | * So some of prechecks for hwpoison (pinning, and testing/setting | |
1857 | * PageHWPoison) should be done in single hugetlb_lock range. | |
1858 | */ | |
1859 | static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) | |
0348d2eb | 1860 | { |
761ad8d7 | 1861 | int res; |
405ce051 NH |
1862 | struct page *p = pfn_to_page(pfn); |
1863 | struct page *head; | |
761ad8d7 | 1864 | unsigned long page_flags; |
e591ef7d | 1865 | bool migratable_cleared = false; |
761ad8d7 | 1866 | |
405ce051 NH |
1867 | *hugetlb = 1; |
1868 | retry: | |
e591ef7d | 1869 | res = get_huge_page_for_hwpoison(pfn, flags, &migratable_cleared); |
405ce051 NH |
1870 | if (res == 2) { /* fallback to normal page handling */ |
1871 | *hugetlb = 0; | |
1872 | return 0; | |
1873 | } else if (res == -EHWPOISON) { | |
96f96763 | 1874 | pr_err("%#lx: already hardware poisoned\n", pfn); |
405ce051 NH |
1875 | if (flags & MF_ACTION_REQUIRED) { |
1876 | head = compound_head(p); | |
a3f5d80e | 1877 | res = kill_accessing_process(current, page_to_pfn(head), flags); |
405ce051 NH |
1878 | } |
1879 | return res; | |
1880 | } else if (res == -EBUSY) { | |
38f6d293 NH |
1881 | if (!(flags & MF_NO_RETRY)) { |
1882 | flags |= MF_NO_RETRY; | |
405ce051 NH |
1883 | goto retry; |
1884 | } | |
b66d00df | 1885 | return action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); |
761ad8d7 NH |
1886 | } |
1887 | ||
405ce051 NH |
1888 | head = compound_head(p); |
1889 | lock_page(head); | |
1890 | ||
1891 | if (hwpoison_filter(p)) { | |
161df60e | 1892 | hugetlb_clear_page_hwpoison(head); |
e591ef7d NH |
1893 | if (migratable_cleared) |
1894 | SetHPageMigratable(head); | |
f36a5543 ML |
1895 | unlock_page(head); |
1896 | if (res == 1) | |
1897 | put_page(head); | |
1898 | return -EOPNOTSUPP; | |
405ce051 NH |
1899 | } |
1900 | ||
405ce051 NH |
1901 | /* |
1902 | * Handling free hugepage. The possible race with hugepage allocation | |
1903 | * or demotion can be prevented by PageHWPoison flag. | |
1904 | */ | |
1905 | if (res == 0) { | |
1906 | unlock_page(head); | |
ceaf8fbe | 1907 | if (__page_handle_poison(p) >= 0) { |
405ce051 NH |
1908 | page_ref_inc(p); |
1909 | res = MF_RECOVERED; | |
ceaf8fbe NH |
1910 | } else { |
1911 | res = MF_FAILED; | |
761ad8d7 | 1912 | } |
b66d00df | 1913 | return action_result(pfn, MF_MSG_FREE_HUGE, res); |
761ad8d7 NH |
1914 | } |
1915 | ||
761ad8d7 NH |
1916 | page_flags = head->flags; |
1917 | ||
ed8c2f49 | 1918 | if (!hwpoison_user_mappings(p, pfn, flags, head)) { |
b66d00df KW |
1919 | unlock_page(head); |
1920 | return action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); | |
761ad8d7 NH |
1921 | } |
1922 | ||
ea6d0630 | 1923 | return identify_page_state(pfn, p, page_flags); |
761ad8d7 | 1924 | } |
00cc790e | 1925 | |
405ce051 NH |
1926 | #else |
1927 | static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) | |
1928 | { | |
1929 | return 0; | |
1930 | } | |
00cc790e | 1931 | |
ac5fcde0 NH |
1932 | static inline unsigned long free_raw_hwp_pages(struct page *hpage, bool flag) |
1933 | { | |
1934 | return 0; | |
1935 | } | |
00cc790e | 1936 | #endif /* CONFIG_HUGETLB_PAGE */ |
761ad8d7 | 1937 | |
b5f1fc98 KW |
1938 | /* Drop the extra refcount in case we come from madvise() */ |
1939 | static void put_ref_page(unsigned long pfn, int flags) | |
1940 | { | |
1941 | struct page *page; | |
1942 | ||
1943 | if (!(flags & MF_COUNT_INCREASED)) | |
1944 | return; | |
1945 | ||
1946 | page = pfn_to_page(pfn); | |
1947 | if (page) | |
1948 | put_page(page); | |
1949 | } | |
1950 | ||
6100e34b DW |
1951 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
1952 | struct dev_pagemap *pgmap) | |
1953 | { | |
00cc790e | 1954 | int rc = -ENXIO; |
6100e34b | 1955 | |
b5f1fc98 | 1956 | put_ref_page(pfn, flags); |
1e8aaedb | 1957 | |
34dc45be | 1958 | /* device metadata space is not recoverable */ |
00cc790e | 1959 | if (!pgmap_pfn_valid(pgmap, pfn)) |
34dc45be | 1960 | goto out; |
61e28cf0 | 1961 | |
6100e34b | 1962 | /* |
33a8f7f2 SR |
1963 | * Call driver's implementation to handle the memory failure, otherwise |
1964 | * fall back to generic handler. | |
6100e34b | 1965 | */ |
65d3440e | 1966 | if (pgmap_has_memory_failure(pgmap)) { |
33a8f7f2 | 1967 | rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags); |
6100e34b | 1968 | /* |
33a8f7f2 SR |
1969 | * Fall back to generic handler too if operation is not |
1970 | * supported inside the driver/device/filesystem. | |
6100e34b | 1971 | */ |
33a8f7f2 SR |
1972 | if (rc != -EOPNOTSUPP) |
1973 | goto out; | |
6100e34b DW |
1974 | } |
1975 | ||
00cc790e | 1976 | rc = mf_generic_kill_procs(pfn, flags, pgmap); |
6100e34b DW |
1977 | out: |
1978 | /* drop pgmap ref acquired in caller */ | |
1979 | put_dev_pagemap(pgmap); | |
1980 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
1981 | return rc; | |
1982 | } | |
1983 | ||
91d00547 NH |
1984 | static DEFINE_MUTEX(mf_mutex); |
1985 | ||
cd42f4a3 TL |
1986 | /** |
1987 | * memory_failure - Handle memory failure of a page. | |
1988 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
1989 | * @flags: fine tune action taken |
1990 | * | |
1991 | * This function is called by the low level machine check code | |
1992 | * of an architecture when it detects hardware memory corruption | |
1993 | * of a page. It tries its best to recover, which includes | |
1994 | * dropping pages, killing processes etc. | |
1995 | * | |
1996 | * The function is primarily of use for corruptions that | |
1997 | * happen outside the current execution context (e.g. when | |
1998 | * detected by a background scrubber) | |
1999 | * | |
2000 | * Must run in process context (e.g. a work queue) with interrupts | |
2001 | * enabled and no spinlocks hold. | |
d1fe111f | 2002 | * |
2003 | * Return: 0 for successfully handled the memory error, | |
9113eaf3 | 2004 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event, |
d1fe111f | 2005 | * < 0(except -EOPNOTSUPP) on failure. |
cd42f4a3 | 2006 | */ |
83b57531 | 2007 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 2008 | { |
6a46079c | 2009 | struct page *p; |
7af446a8 | 2010 | struct page *hpage; |
6100e34b | 2011 | struct dev_pagemap *pgmap; |
171936dd | 2012 | int res = 0; |
524fca1e | 2013 | unsigned long page_flags; |
a8b2c2ce | 2014 | bool retry = true; |
405ce051 | 2015 | int hugetlb = 0; |
6a46079c AK |
2016 | |
2017 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 2018 | panic("Memory failure on page %lx", pfn); |
6a46079c | 2019 | |
03b122da TL |
2020 | mutex_lock(&mf_mutex); |
2021 | ||
67f22ba7 | 2022 | if (!(flags & MF_SW_SIMULATED)) |
2023 | hw_memory_failure = true; | |
2024 | ||
96c804a6 DH |
2025 | p = pfn_to_online_page(pfn); |
2026 | if (!p) { | |
03b122da TL |
2027 | res = arch_memory_failure(pfn, flags); |
2028 | if (res == 0) | |
2029 | goto unlock_mutex; | |
2030 | ||
96c804a6 DH |
2031 | if (pfn_valid(pfn)) { |
2032 | pgmap = get_dev_pagemap(pfn, NULL); | |
03b122da TL |
2033 | if (pgmap) { |
2034 | res = memory_failure_dev_pagemap(pfn, flags, | |
2035 | pgmap); | |
2036 | goto unlock_mutex; | |
2037 | } | |
96c804a6 | 2038 | } |
96f96763 | 2039 | pr_err("%#lx: memory outside kernel control\n", pfn); |
03b122da TL |
2040 | res = -ENXIO; |
2041 | goto unlock_mutex; | |
6a46079c AK |
2042 | } |
2043 | ||
a8b2c2ce | 2044 | try_again: |
405ce051 NH |
2045 | res = try_memory_failure_hugetlb(pfn, flags, &hugetlb); |
2046 | if (hugetlb) | |
171936dd | 2047 | goto unlock_mutex; |
171936dd | 2048 | |
6a46079c | 2049 | if (TestSetPageHWPoison(p)) { |
96f96763 | 2050 | pr_err("%#lx: already hardware poisoned\n", pfn); |
47af12ba | 2051 | res = -EHWPOISON; |
a3f5d80e NH |
2052 | if (flags & MF_ACTION_REQUIRED) |
2053 | res = kill_accessing_process(current, pfn, flags); | |
f361e246 NH |
2054 | if (flags & MF_COUNT_INCREASED) |
2055 | put_page(p); | |
171936dd | 2056 | goto unlock_mutex; |
6a46079c AK |
2057 | } |
2058 | ||
75ee64b3 | 2059 | hpage = compound_head(p); |
6a46079c AK |
2060 | |
2061 | /* | |
2062 | * We need/can do nothing about count=0 pages. | |
2063 | * 1) it's a free page, and therefore in safe hand: | |
9cf28191 | 2064 | * check_new_page() will be the gate keeper. |
761ad8d7 | 2065 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
2066 | * Implies some kernel user: cannot stop them from |
2067 | * R/W the page; let's pray that the page has been | |
2068 | * used and will be freed some time later. | |
2069 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 2070 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 2071 | */ |
0ed950d1 NH |
2072 | if (!(flags & MF_COUNT_INCREASED)) { |
2073 | res = get_hwpoison_page(p, flags); | |
2074 | if (!res) { | |
2075 | if (is_free_buddy_page(p)) { | |
2076 | if (take_page_off_buddy(p)) { | |
2077 | page_ref_inc(p); | |
2078 | res = MF_RECOVERED; | |
2079 | } else { | |
2080 | /* We lost the race, try again */ | |
2081 | if (retry) { | |
2082 | ClearPageHWPoison(p); | |
0ed950d1 NH |
2083 | retry = false; |
2084 | goto try_again; | |
2085 | } | |
2086 | res = MF_FAILED; | |
a8b2c2ce | 2087 | } |
b66d00df | 2088 | res = action_result(pfn, MF_MSG_BUDDY, res); |
0ed950d1 | 2089 | } else { |
b66d00df | 2090 | res = action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); |
a8b2c2ce | 2091 | } |
0ed950d1 NH |
2092 | goto unlock_mutex; |
2093 | } else if (res < 0) { | |
b66d00df | 2094 | res = action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); |
0ed950d1 | 2095 | goto unlock_mutex; |
8d22ba1b | 2096 | } |
6a46079c AK |
2097 | } |
2098 | ||
761ad8d7 | 2099 | if (PageTransHuge(hpage)) { |
eac96c3e YS |
2100 | /* |
2101 | * The flag must be set after the refcount is bumped | |
2102 | * otherwise it may race with THP split. | |
2103 | * And the flag can't be set in get_hwpoison_page() since | |
2104 | * it is called by soft offline too and it is just called | |
2105 | * for !MF_COUNT_INCREASE. So here seems to be the best | |
2106 | * place. | |
2107 | * | |
2108 | * Don't need care about the above error handling paths for | |
2109 | * get_hwpoison_page() since they handle either free page | |
2110 | * or unhandlable page. The refcount is bumped iff the | |
2111 | * page is a valid handlable page. | |
2112 | */ | |
2113 | SetPageHasHWPoisoned(hpage); | |
2ace36f0 | 2114 | if (try_to_split_thp_page(p) < 0) { |
b66d00df | 2115 | res = action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED); |
171936dd | 2116 | goto unlock_mutex; |
5d1fd5dc | 2117 | } |
415c64c1 | 2118 | VM_BUG_ON_PAGE(!page_count(p), p); |
415c64c1 NH |
2119 | } |
2120 | ||
e43c3afb WF |
2121 | /* |
2122 | * We ignore non-LRU pages for good reasons. | |
2123 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 2124 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
2125 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
2126 | * The check (unnecessarily) ignores LRU pages being isolated and | |
2127 | * walked by the page reclaim code, however that's not a big loss. | |
2128 | */ | |
d0505e9f | 2129 | shake_page(p); |
e43c3afb | 2130 | |
761ad8d7 | 2131 | lock_page(p); |
847ce401 | 2132 | |
f37d4298 | 2133 | /* |
75ee64b3 ML |
2134 | * We're only intended to deal with the non-Compound page here. |
2135 | * However, the page could have changed compound pages due to | |
2136 | * race window. If this happens, we could try again to hopefully | |
2137 | * handle the page next round. | |
f37d4298 | 2138 | */ |
75ee64b3 ML |
2139 | if (PageCompound(p)) { |
2140 | if (retry) { | |
e240ac52 | 2141 | ClearPageHWPoison(p); |
75ee64b3 ML |
2142 | unlock_page(p); |
2143 | put_page(p); | |
2144 | flags &= ~MF_COUNT_INCREASED; | |
2145 | retry = false; | |
2146 | goto try_again; | |
2147 | } | |
b66d00df | 2148 | res = action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
171936dd | 2149 | goto unlock_page; |
f37d4298 AK |
2150 | } |
2151 | ||
524fca1e NH |
2152 | /* |
2153 | * We use page flags to determine what action should be taken, but | |
2154 | * the flags can be modified by the error containment action. One | |
2155 | * example is an mlocked page, where PG_mlocked is cleared by | |
2156 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
2157 | * correctly, we save a copy of the page flags at this time. | |
2158 | */ | |
7d9d46ac | 2159 | page_flags = p->flags; |
524fca1e | 2160 | |
7c116f2b | 2161 | if (hwpoison_filter(p)) { |
2fe62e22 | 2162 | ClearPageHWPoison(p); |
761ad8d7 | 2163 | unlock_page(p); |
dd6e2402 | 2164 | put_page(p); |
d1fe111f | 2165 | res = -EOPNOTSUPP; |
171936dd | 2166 | goto unlock_mutex; |
7c116f2b | 2167 | } |
847ce401 | 2168 | |
e8675d29 | 2169 | /* |
2170 | * __munlock_pagevec may clear a writeback page's LRU flag without | |
2171 | * page_lock. We need wait writeback completion for this page or it | |
2172 | * may trigger vfs BUG while evict inode. | |
2173 | */ | |
b04d3eeb | 2174 | if (!PageLRU(p) && !PageWriteback(p)) |
0bc1f8b0 CY |
2175 | goto identify_page_state; |
2176 | ||
6edd6cc6 NH |
2177 | /* |
2178 | * It's very difficult to mess with pages currently under IO | |
2179 | * and in many cases impossible, so we just avoid it here. | |
2180 | */ | |
6a46079c AK |
2181 | wait_on_page_writeback(p); |
2182 | ||
2183 | /* | |
2184 | * Now take care of user space mappings. | |
6ffcd825 | 2185 | * Abort on fail: __filemap_remove_folio() assumes unmapped page. |
6a46079c | 2186 | */ |
ed8c2f49 | 2187 | if (!hwpoison_user_mappings(p, pfn, flags, p)) { |
b66d00df | 2188 | res = action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
171936dd | 2189 | goto unlock_page; |
1668bfd5 | 2190 | } |
6a46079c AK |
2191 | |
2192 | /* | |
2193 | * Torn down by someone else? | |
2194 | */ | |
dc2a1cbf | 2195 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
b66d00df | 2196 | res = action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
171936dd | 2197 | goto unlock_page; |
6a46079c AK |
2198 | } |
2199 | ||
0bc1f8b0 | 2200 | identify_page_state: |
0348d2eb | 2201 | res = identify_page_state(pfn, p, page_flags); |
ea6d0630 NH |
2202 | mutex_unlock(&mf_mutex); |
2203 | return res; | |
171936dd | 2204 | unlock_page: |
761ad8d7 | 2205 | unlock_page(p); |
171936dd TL |
2206 | unlock_mutex: |
2207 | mutex_unlock(&mf_mutex); | |
6a46079c AK |
2208 | return res; |
2209 | } | |
cd42f4a3 | 2210 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 2211 | |
ea8f5fb8 HY |
2212 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
2213 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
2214 | ||
2215 | struct memory_failure_entry { | |
2216 | unsigned long pfn; | |
ea8f5fb8 HY |
2217 | int flags; |
2218 | }; | |
2219 | ||
2220 | struct memory_failure_cpu { | |
2221 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
2222 | MEMORY_FAILURE_FIFO_SIZE); | |
2223 | spinlock_t lock; | |
2224 | struct work_struct work; | |
2225 | }; | |
2226 | ||
2227 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
2228 | ||
2229 | /** | |
2230 | * memory_failure_queue - Schedule handling memory failure of a page. | |
2231 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
2232 | * @flags: Flags for memory failure handling |
2233 | * | |
2234 | * This function is called by the low level hardware error handler | |
2235 | * when it detects hardware memory corruption of a page. It schedules | |
2236 | * the recovering of error page, including dropping pages, killing | |
2237 | * processes etc. | |
2238 | * | |
2239 | * The function is primarily of use for corruptions that | |
2240 | * happen outside the current execution context (e.g. when | |
2241 | * detected by a background scrubber) | |
2242 | * | |
2243 | * Can run in IRQ context. | |
2244 | */ | |
83b57531 | 2245 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
2246 | { |
2247 | struct memory_failure_cpu *mf_cpu; | |
2248 | unsigned long proc_flags; | |
2249 | struct memory_failure_entry entry = { | |
2250 | .pfn = pfn, | |
ea8f5fb8 HY |
2251 | .flags = flags, |
2252 | }; | |
2253 | ||
2254 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
2255 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 2256 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
2257 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
2258 | else | |
96f96763 | 2259 | pr_err("buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
2260 | pfn); |
2261 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
2262 | put_cpu_var(memory_failure_cpu); | |
2263 | } | |
2264 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
2265 | ||
2266 | static void memory_failure_work_func(struct work_struct *work) | |
2267 | { | |
2268 | struct memory_failure_cpu *mf_cpu; | |
2269 | struct memory_failure_entry entry = { 0, }; | |
2270 | unsigned long proc_flags; | |
2271 | int gotten; | |
2272 | ||
06202231 | 2273 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
ea8f5fb8 HY |
2274 | for (;;) { |
2275 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
2276 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
2277 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
2278 | if (!gotten) | |
2279 | break; | |
cf870c70 | 2280 | if (entry.flags & MF_SOFT_OFFLINE) |
feec24a6 | 2281 | soft_offline_page(entry.pfn, entry.flags); |
cf870c70 | 2282 | else |
83b57531 | 2283 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
2284 | } |
2285 | } | |
2286 | ||
06202231 JM |
2287 | /* |
2288 | * Process memory_failure work queued on the specified CPU. | |
2289 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. | |
2290 | */ | |
2291 | void memory_failure_queue_kick(int cpu) | |
2292 | { | |
2293 | struct memory_failure_cpu *mf_cpu; | |
2294 | ||
2295 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
2296 | cancel_work_sync(&mf_cpu->work); | |
2297 | memory_failure_work_func(&mf_cpu->work); | |
2298 | } | |
2299 | ||
ea8f5fb8 HY |
2300 | static int __init memory_failure_init(void) |
2301 | { | |
2302 | struct memory_failure_cpu *mf_cpu; | |
2303 | int cpu; | |
2304 | ||
2305 | for_each_possible_cpu(cpu) { | |
2306 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
2307 | spin_lock_init(&mf_cpu->lock); | |
2308 | INIT_KFIFO(mf_cpu->fifo); | |
2309 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
2310 | } | |
2311 | ||
2312 | return 0; | |
2313 | } | |
2314 | core_initcall(memory_failure_init); | |
2315 | ||
96f96763 KW |
2316 | #undef pr_fmt |
2317 | #define pr_fmt(fmt) "" fmt | |
a5f65109 NH |
2318 | #define unpoison_pr_info(fmt, pfn, rs) \ |
2319 | ({ \ | |
2320 | if (__ratelimit(rs)) \ | |
2321 | pr_info(fmt, pfn); \ | |
2322 | }) | |
2323 | ||
847ce401 WF |
2324 | /** |
2325 | * unpoison_memory - Unpoison a previously poisoned page | |
2326 | * @pfn: Page number of the to be unpoisoned page | |
2327 | * | |
2328 | * Software-unpoison a page that has been poisoned by | |
2329 | * memory_failure() earlier. | |
2330 | * | |
2331 | * This is only done on the software-level, so it only works | |
2332 | * for linux injected failures, not real hardware failures | |
2333 | * | |
2334 | * Returns 0 for success, otherwise -errno. | |
2335 | */ | |
2336 | int unpoison_memory(unsigned long pfn) | |
2337 | { | |
2338 | struct page *page; | |
2339 | struct page *p; | |
bf181c58 | 2340 | int ret = -EBUSY; |
ac5fcde0 | 2341 | unsigned long count = 1; |
5033091d | 2342 | bool huge = false; |
a5f65109 NH |
2343 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
2344 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
2345 | |
2346 | if (!pfn_valid(pfn)) | |
2347 | return -ENXIO; | |
2348 | ||
2349 | p = pfn_to_page(pfn); | |
2350 | page = compound_head(p); | |
2351 | ||
91d00547 NH |
2352 | mutex_lock(&mf_mutex); |
2353 | ||
67f22ba7 | 2354 | if (hw_memory_failure) { |
2355 | unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n", | |
2356 | pfn, &unpoison_rs); | |
2357 | ret = -EOPNOTSUPP; | |
2358 | goto unlock_mutex; | |
2359 | } | |
2360 | ||
847ce401 | 2361 | if (!PageHWPoison(p)) { |
495367c0 | 2362 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 2363 | pfn, &unpoison_rs); |
91d00547 | 2364 | goto unlock_mutex; |
847ce401 WF |
2365 | } |
2366 | ||
230ac719 | 2367 | if (page_count(page) > 1) { |
495367c0 | 2368 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 2369 | pfn, &unpoison_rs); |
91d00547 | 2370 | goto unlock_mutex; |
230ac719 NH |
2371 | } |
2372 | ||
2373 | if (page_mapped(page)) { | |
495367c0 | 2374 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 2375 | pfn, &unpoison_rs); |
91d00547 | 2376 | goto unlock_mutex; |
230ac719 NH |
2377 | } |
2378 | ||
2379 | if (page_mapping(page)) { | |
495367c0 | 2380 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 2381 | pfn, &unpoison_rs); |
91d00547 | 2382 | goto unlock_mutex; |
0cea3fdc WL |
2383 | } |
2384 | ||
e9ff3ba7 | 2385 | if (PageSlab(page) || PageTable(page) || PageReserved(page)) |
91d00547 | 2386 | goto unlock_mutex; |
847ce401 | 2387 | |
bf181c58 NH |
2388 | ret = get_hwpoison_page(p, MF_UNPOISON); |
2389 | if (!ret) { | |
ac5fcde0 | 2390 | if (PageHuge(p)) { |
5033091d | 2391 | huge = true; |
ac5fcde0 NH |
2392 | count = free_raw_hwp_pages(page, false); |
2393 | if (count == 0) { | |
2394 | ret = -EBUSY; | |
2395 | goto unlock_mutex; | |
2396 | } | |
2397 | } | |
c8bd84f7 | 2398 | ret = TestClearPageHWPoison(page) ? 0 : -EBUSY; |
bf181c58 NH |
2399 | } else if (ret < 0) { |
2400 | if (ret == -EHWPOISON) { | |
c8bd84f7 | 2401 | ret = put_page_back_buddy(p) ? 0 : -EBUSY; |
bf181c58 NH |
2402 | } else |
2403 | unpoison_pr_info("Unpoison: failed to grab page %#lx\n", | |
2404 | pfn, &unpoison_rs); | |
2405 | } else { | |
ac5fcde0 | 2406 | if (PageHuge(p)) { |
5033091d | 2407 | huge = true; |
ac5fcde0 NH |
2408 | count = free_raw_hwp_pages(page, false); |
2409 | if (count == 0) { | |
2410 | ret = -EBUSY; | |
6bbabd04 | 2411 | put_page(page); |
ac5fcde0 NH |
2412 | goto unlock_mutex; |
2413 | } | |
2414 | } | |
847ce401 | 2415 | |
dd6e2402 | 2416 | put_page(page); |
e0ff4280 | 2417 | if (TestClearPageHWPoison(p)) { |
bf181c58 NH |
2418 | put_page(page); |
2419 | ret = 0; | |
2420 | } | |
2421 | } | |
847ce401 | 2422 | |
91d00547 NH |
2423 | unlock_mutex: |
2424 | mutex_unlock(&mf_mutex); | |
e0ff4280 | 2425 | if (!ret) { |
5033091d NH |
2426 | if (!huge) |
2427 | num_poisoned_pages_sub(pfn, 1); | |
c8bd84f7 | 2428 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", |
2429 | page_to_pfn(p), &unpoison_rs); | |
2430 | } | |
91d00547 | 2431 | return ret; |
847ce401 WF |
2432 | } |
2433 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 2434 | |
6b9a217e | 2435 | static bool isolate_page(struct page *page, struct list_head *pagelist) |
d950b958 | 2436 | { |
6b9a217e | 2437 | bool isolated = false; |
d950b958 | 2438 | |
6b9a217e | 2439 | if (PageHuge(page)) { |
7ce82f4c | 2440 | isolated = !isolate_hugetlb(page, pagelist); |
6b9a217e | 2441 | } else { |
da294991 ML |
2442 | bool lru = !__PageMovable(page); |
2443 | ||
6b9a217e OS |
2444 | if (lru) |
2445 | isolated = !isolate_lru_page(page); | |
2446 | else | |
da294991 ML |
2447 | isolated = !isolate_movable_page(page, |
2448 | ISOLATE_UNEVICTABLE); | |
6b9a217e | 2449 | |
da294991 | 2450 | if (isolated) { |
6b9a217e | 2451 | list_add(&page->lru, pagelist); |
da294991 ML |
2452 | if (lru) |
2453 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
2454 | page_is_file_lru(page)); | |
2455 | } | |
0ebff32c | 2456 | } |
d950b958 | 2457 | |
03613808 | 2458 | /* |
6b9a217e OS |
2459 | * If we succeed to isolate the page, we grabbed another refcount on |
2460 | * the page, so we can safely drop the one we got from get_any_pages(). | |
2461 | * If we failed to isolate the page, it means that we cannot go further | |
2462 | * and we will return an error, so drop the reference we got from | |
2463 | * get_any_pages() as well. | |
03613808 | 2464 | */ |
6b9a217e OS |
2465 | put_page(page); |
2466 | return isolated; | |
d950b958 NH |
2467 | } |
2468 | ||
6b9a217e | 2469 | /* |
48309e1f | 2470 | * soft_offline_in_use_page handles hugetlb-pages and non-hugetlb pages. |
6b9a217e OS |
2471 | * If the page is a non-dirty unmapped page-cache page, it simply invalidates. |
2472 | * If the page is mapped, it migrates the contents over. | |
2473 | */ | |
48309e1f | 2474 | static int soft_offline_in_use_page(struct page *page) |
af8fae7c | 2475 | { |
d6c75dc2 | 2476 | long ret = 0; |
af8fae7c | 2477 | unsigned long pfn = page_to_pfn(page); |
6b9a217e OS |
2478 | struct page *hpage = compound_head(page); |
2479 | char const *msg_page[] = {"page", "hugepage"}; | |
2480 | bool huge = PageHuge(page); | |
2481 | LIST_HEAD(pagelist); | |
54608759 JK |
2482 | struct migration_target_control mtc = { |
2483 | .nid = NUMA_NO_NODE, | |
2484 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
2485 | }; | |
facb6011 | 2486 | |
48309e1f KW |
2487 | if (!huge && PageTransHuge(hpage)) { |
2488 | if (try_to_split_thp_page(page)) { | |
2489 | pr_info("soft offline: %#lx: thp split failed\n", pfn); | |
2490 | return -EBUSY; | |
2491 | } | |
2492 | hpage = page; | |
2493 | } | |
2494 | ||
0ebff32c | 2495 | lock_page(page); |
6b9a217e OS |
2496 | if (!PageHuge(page)) |
2497 | wait_on_page_writeback(page); | |
af8fae7c NH |
2498 | if (PageHWPoison(page)) { |
2499 | unlock_page(page); | |
dd6e2402 | 2500 | put_page(page); |
af8fae7c | 2501 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
5a2ffca3 | 2502 | return 0; |
af8fae7c | 2503 | } |
6b9a217e | 2504 | |
593396b8 | 2505 | if (!PageHuge(page) && PageLRU(page) && !PageSwapCache(page)) |
6b9a217e OS |
2506 | /* |
2507 | * Try to invalidate first. This should work for | |
2508 | * non dirty unmapped page cache pages. | |
2509 | */ | |
2510 | ret = invalidate_inode_page(page); | |
facb6011 | 2511 | unlock_page(page); |
6b9a217e | 2512 | |
6b9a217e | 2513 | if (ret) { |
fb46e735 | 2514 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
6b9a217e | 2515 | page_handle_poison(page, false, true); |
af8fae7c | 2516 | return 0; |
facb6011 AK |
2517 | } |
2518 | ||
6b9a217e | 2519 | if (isolate_page(hpage, &pagelist)) { |
54608759 | 2520 | ret = migrate_pages(&pagelist, alloc_migration_target, NULL, |
5ac95884 | 2521 | (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL); |
79f5f8fa | 2522 | if (!ret) { |
6b9a217e OS |
2523 | bool release = !huge; |
2524 | ||
2525 | if (!page_handle_poison(page, huge, release)) | |
2526 | ret = -EBUSY; | |
79f5f8fa | 2527 | } else { |
85fbe5d1 YX |
2528 | if (!list_empty(&pagelist)) |
2529 | putback_movable_pages(&pagelist); | |
59c82b70 | 2530 | |
d6c75dc2 | 2531 | pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n", |
23efd080 | 2532 | pfn, msg_page[huge], ret, &page->flags); |
facb6011 | 2533 | if (ret > 0) |
3f4b815a | 2534 | ret = -EBUSY; |
facb6011 AK |
2535 | } |
2536 | } else { | |
23efd080 MWO |
2537 | pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n", |
2538 | pfn, msg_page[huge], page_count(page), &page->flags); | |
6b9a217e | 2539 | ret = -EBUSY; |
facb6011 | 2540 | } |
facb6011 AK |
2541 | return ret; |
2542 | } | |
86e05773 WL |
2543 | |
2544 | /** | |
2545 | * soft_offline_page - Soft offline a page. | |
feec24a6 | 2546 | * @pfn: pfn to soft-offline |
86e05773 WL |
2547 | * @flags: flags. Same as memory_failure(). |
2548 | * | |
9113eaf3 | 2549 | * Returns 0 on success |
2550 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event | |
2551 | * < 0 otherwise negated errno. | |
86e05773 WL |
2552 | * |
2553 | * Soft offline a page, by migration or invalidation, | |
2554 | * without killing anything. This is for the case when | |
2555 | * a page is not corrupted yet (so it's still valid to access), | |
2556 | * but has had a number of corrected errors and is better taken | |
2557 | * out. | |
2558 | * | |
2559 | * The actual policy on when to do that is maintained by | |
2560 | * user space. | |
2561 | * | |
2562 | * This should never impact any application or cause data loss, | |
2563 | * however it might take some time. | |
2564 | * | |
2565 | * This is not a 100% solution for all memory, but tries to be | |
2566 | * ``good enough'' for the majority of memory. | |
2567 | */ | |
feec24a6 | 2568 | int soft_offline_page(unsigned long pfn, int flags) |
86e05773 WL |
2569 | { |
2570 | int ret; | |
b94e0282 | 2571 | bool try_again = true; |
b5f1fc98 | 2572 | struct page *page; |
dad4e5b3 | 2573 | |
183a7c5d KW |
2574 | if (!pfn_valid(pfn)) { |
2575 | WARN_ON_ONCE(flags & MF_COUNT_INCREASED); | |
feec24a6 | 2576 | return -ENXIO; |
183a7c5d | 2577 | } |
dad4e5b3 | 2578 | |
feec24a6 NH |
2579 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ |
2580 | page = pfn_to_online_page(pfn); | |
dad4e5b3 | 2581 | if (!page) { |
b5f1fc98 | 2582 | put_ref_page(pfn, flags); |
86a66810 | 2583 | return -EIO; |
dad4e5b3 | 2584 | } |
86a66810 | 2585 | |
91d00547 NH |
2586 | mutex_lock(&mf_mutex); |
2587 | ||
86e05773 | 2588 | if (PageHWPoison(page)) { |
8295d535 | 2589 | pr_info("%s: %#lx page already poisoned\n", __func__, pfn); |
b5f1fc98 | 2590 | put_ref_page(pfn, flags); |
91d00547 | 2591 | mutex_unlock(&mf_mutex); |
5a2ffca3 | 2592 | return 0; |
86e05773 | 2593 | } |
86e05773 | 2594 | |
b94e0282 | 2595 | retry: |
bfc8c901 | 2596 | get_online_mems(); |
bf6445bc | 2597 | ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE); |
bfc8c901 | 2598 | put_online_mems(); |
4e41a30c | 2599 | |
9113eaf3 | 2600 | if (hwpoison_filter(page)) { |
2601 | if (ret > 0) | |
2602 | put_page(page); | |
9113eaf3 | 2603 | |
2604 | mutex_unlock(&mf_mutex); | |
2605 | return -EOPNOTSUPP; | |
2606 | } | |
2607 | ||
8295d535 | 2608 | if (ret > 0) { |
6b9a217e | 2609 | ret = soft_offline_in_use_page(page); |
8295d535 | 2610 | } else if (ret == 0) { |
7adb4588 | 2611 | if (!page_handle_poison(page, true, false) && try_again) { |
b94e0282 | 2612 | try_again = false; |
2a57d83c | 2613 | flags &= ~MF_COUNT_INCREASED; |
b94e0282 OS |
2614 | goto retry; |
2615 | } | |
8295d535 | 2616 | } |
4e41a30c | 2617 | |
91d00547 NH |
2618 | mutex_unlock(&mf_mutex); |
2619 | ||
86e05773 WL |
2620 | return ret; |
2621 | } |