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. |
c33c7948 | 9 | * |
1c80b990 AK |
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 | |
c33c7948 RR |
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 | |
1c80b990 | 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. |
c33c7948 | 28 | * |
1c80b990 | 29 | * There are several operations here with exponential complexity because |
c33c7948 RR |
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 | |
1c80b990 | 32 | * has non linear complexity with the number. But since memory corruptions |
c33c7948 | 33 | * are rare we hope to get away with this. This avoids impacting the core |
1c80b990 | 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> | |
3f07c014 | 42 | #include <linux/sched/signal.h> |
29930025 | 43 | #include <linux/sched/task.h> |
96c84dde | 44 | #include <linux/dax.h> |
01e00f88 | 45 | #include <linux/ksm.h> |
6a46079c | 46 | #include <linux/rmap.h> |
b9e15baf | 47 | #include <linux/export.h> |
6a46079c AK |
48 | #include <linux/pagemap.h> |
49 | #include <linux/swap.h> | |
50 | #include <linux/backing-dev.h> | |
facb6011 | 51 | #include <linux/migrate.h> |
5a0e3ad6 | 52 | #include <linux/slab.h> |
bf998156 | 53 | #include <linux/swapops.h> |
7af446a8 | 54 | #include <linux/hugetlb.h> |
20d6c96b | 55 | #include <linux/memory_hotplug.h> |
5db8a73a | 56 | #include <linux/mm_inline.h> |
6100e34b | 57 | #include <linux/memremap.h> |
ea8f5fb8 | 58 | #include <linux/kfifo.h> |
a5f65109 | 59 | #include <linux/ratelimit.h> |
a3f5d80e | 60 | #include <linux/pagewalk.h> |
a7605426 | 61 | #include <linux/shmem_fs.h> |
8cbc82f3 | 62 | #include <linux/sysctl.h> |
014bb1de | 63 | #include "swap.h" |
6a46079c | 64 | #include "internal.h" |
97f0b134 | 65 | #include "ras/ras_event.h" |
6a46079c | 66 | |
8cbc82f3 | 67 | static int sysctl_memory_failure_early_kill __read_mostly; |
6a46079c | 68 | |
8cbc82f3 | 69 | static int sysctl_memory_failure_recovery __read_mostly = 1; |
6a46079c | 70 | |
293c07e3 | 71 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 72 | |
67f22ba7 | 73 | static bool hw_memory_failure __read_mostly = false; |
74 | ||
b79f8eb4 JY |
75 | static DEFINE_MUTEX(mf_mutex); |
76 | ||
1a7d018d | 77 | 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 | ||
1a7d018d | 83 | 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 | ||
44b8f8bf JY |
90 | /** |
91 | * MF_ATTR_RO - Create sysfs entry for each memory failure statistics. | |
92 | * @_name: name of the file in the per NUMA sysfs directory. | |
93 | */ | |
94 | #define MF_ATTR_RO(_name) \ | |
95 | static ssize_t _name##_show(struct device *dev, \ | |
96 | struct device_attribute *attr, \ | |
97 | char *buf) \ | |
98 | { \ | |
99 | struct memory_failure_stats *mf_stats = \ | |
100 | &NODE_DATA(dev->id)->mf_stats; \ | |
101 | return sprintf(buf, "%lu\n", mf_stats->_name); \ | |
102 | } \ | |
103 | static DEVICE_ATTR_RO(_name) | |
104 | ||
105 | MF_ATTR_RO(total); | |
106 | MF_ATTR_RO(ignored); | |
107 | MF_ATTR_RO(failed); | |
108 | MF_ATTR_RO(delayed); | |
109 | MF_ATTR_RO(recovered); | |
110 | ||
111 | static struct attribute *memory_failure_attr[] = { | |
112 | &dev_attr_total.attr, | |
113 | &dev_attr_ignored.attr, | |
114 | &dev_attr_failed.attr, | |
115 | &dev_attr_delayed.attr, | |
116 | &dev_attr_recovered.attr, | |
117 | NULL, | |
118 | }; | |
119 | ||
120 | const struct attribute_group memory_failure_attr_group = { | |
121 | .name = "memory_failure", | |
122 | .attrs = memory_failure_attr, | |
123 | }; | |
124 | ||
8cbc82f3 KW |
125 | static struct ctl_table memory_failure_table[] = { |
126 | { | |
127 | .procname = "memory_failure_early_kill", | |
128 | .data = &sysctl_memory_failure_early_kill, | |
129 | .maxlen = sizeof(sysctl_memory_failure_early_kill), | |
130 | .mode = 0644, | |
131 | .proc_handler = proc_dointvec_minmax, | |
132 | .extra1 = SYSCTL_ZERO, | |
133 | .extra2 = SYSCTL_ONE, | |
134 | }, | |
135 | { | |
136 | .procname = "memory_failure_recovery", | |
137 | .data = &sysctl_memory_failure_recovery, | |
138 | .maxlen = sizeof(sysctl_memory_failure_recovery), | |
139 | .mode = 0644, | |
140 | .proc_handler = proc_dointvec_minmax, | |
141 | .extra1 = SYSCTL_ZERO, | |
142 | .extra2 = SYSCTL_ONE, | |
143 | }, | |
144 | { } | |
145 | }; | |
146 | ||
7453bf62 NH |
147 | /* |
148 | * Return values: | |
149 | * 1: the page is dissolved (if needed) and taken off from buddy, | |
150 | * 0: the page is dissolved (if needed) and not taken off from buddy, | |
151 | * < 0: failed to dissolve. | |
152 | */ | |
153 | static int __page_handle_poison(struct page *page) | |
510d25c9 | 154 | { |
f87060d3 | 155 | int ret; |
510d25c9 NH |
156 | |
157 | zone_pcp_disable(page_zone(page)); | |
158 | ret = dissolve_free_huge_page(page); | |
159 | if (!ret) | |
160 | ret = take_page_off_buddy(page); | |
161 | zone_pcp_enable(page_zone(page)); | |
162 | ||
7453bf62 | 163 | return ret; |
510d25c9 NH |
164 | } |
165 | ||
6b9a217e | 166 | static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release) |
06be6ff3 | 167 | { |
6b9a217e OS |
168 | if (hugepage_or_freepage) { |
169 | /* | |
170 | * Doing this check for free pages is also fine since dissolve_free_huge_page | |
171 | * returns 0 for non-hugetlb pages as well. | |
172 | */ | |
7453bf62 | 173 | if (__page_handle_poison(page) <= 0) |
6b9a217e OS |
174 | /* |
175 | * We could fail to take off the target page from buddy | |
f0953a1b | 176 | * for example due to racy page allocation, but that's |
6b9a217e OS |
177 | * acceptable because soft-offlined page is not broken |
178 | * and if someone really want to use it, they should | |
179 | * take it. | |
180 | */ | |
181 | return false; | |
182 | } | |
183 | ||
06be6ff3 | 184 | SetPageHWPoison(page); |
79f5f8fa OS |
185 | if (release) |
186 | put_page(page); | |
06be6ff3 | 187 | page_ref_inc(page); |
a46c9304 | 188 | num_poisoned_pages_inc(page_to_pfn(page)); |
6b9a217e OS |
189 | |
190 | return true; | |
06be6ff3 OS |
191 | } |
192 | ||
611b9fd8 | 193 | #if IS_ENABLED(CONFIG_HWPOISON_INJECT) |
27df5068 | 194 | |
1bfe5feb | 195 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
196 | u32 hwpoison_filter_dev_major = ~0U; |
197 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
198 | u64 hwpoison_filter_flags_mask; |
199 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 200 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
201 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
202 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
203 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
204 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
205 | |
206 | static int hwpoison_filter_dev(struct page *p) | |
207 | { | |
208 | struct address_space *mapping; | |
209 | dev_t dev; | |
210 | ||
211 | if (hwpoison_filter_dev_major == ~0U && | |
212 | hwpoison_filter_dev_minor == ~0U) | |
213 | return 0; | |
214 | ||
7c116f2b WF |
215 | mapping = page_mapping(p); |
216 | if (mapping == NULL || mapping->host == NULL) | |
217 | return -EINVAL; | |
218 | ||
219 | dev = mapping->host->i_sb->s_dev; | |
220 | if (hwpoison_filter_dev_major != ~0U && | |
221 | hwpoison_filter_dev_major != MAJOR(dev)) | |
222 | return -EINVAL; | |
223 | if (hwpoison_filter_dev_minor != ~0U && | |
224 | hwpoison_filter_dev_minor != MINOR(dev)) | |
225 | return -EINVAL; | |
226 | ||
227 | return 0; | |
228 | } | |
229 | ||
478c5ffc WF |
230 | static int hwpoison_filter_flags(struct page *p) |
231 | { | |
232 | if (!hwpoison_filter_flags_mask) | |
233 | return 0; | |
234 | ||
235 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
236 | hwpoison_filter_flags_value) | |
237 | return 0; | |
238 | else | |
239 | return -EINVAL; | |
240 | } | |
241 | ||
4fd466eb AK |
242 | /* |
243 | * This allows stress tests to limit test scope to a collection of tasks | |
244 | * by putting them under some memcg. This prevents killing unrelated/important | |
245 | * processes such as /sbin/init. Note that the target task may share clean | |
246 | * pages with init (eg. libc text), which is harmless. If the target task | |
247 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
248 | * is also included in the memcg. At last, due to race conditions this filter | |
249 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
250 | * a freed page. | |
251 | */ | |
94a59fb3 | 252 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
253 | u64 hwpoison_filter_memcg; |
254 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
255 | static int hwpoison_filter_task(struct page *p) | |
256 | { | |
4fd466eb AK |
257 | if (!hwpoison_filter_memcg) |
258 | return 0; | |
259 | ||
94a59fb3 | 260 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
261 | return -EINVAL; |
262 | ||
263 | return 0; | |
264 | } | |
265 | #else | |
266 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
267 | #endif | |
268 | ||
7c116f2b WF |
269 | int hwpoison_filter(struct page *p) |
270 | { | |
1bfe5feb HL |
271 | if (!hwpoison_filter_enable) |
272 | return 0; | |
273 | ||
7c116f2b WF |
274 | if (hwpoison_filter_dev(p)) |
275 | return -EINVAL; | |
276 | ||
478c5ffc WF |
277 | if (hwpoison_filter_flags(p)) |
278 | return -EINVAL; | |
279 | ||
4fd466eb AK |
280 | if (hwpoison_filter_task(p)) |
281 | return -EINVAL; | |
282 | ||
7c116f2b WF |
283 | return 0; |
284 | } | |
27df5068 AK |
285 | #else |
286 | int hwpoison_filter(struct page *p) | |
287 | { | |
288 | return 0; | |
289 | } | |
290 | #endif | |
291 | ||
7c116f2b WF |
292 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
293 | ||
ae1139ec DW |
294 | /* |
295 | * Kill all processes that have a poisoned page mapped and then isolate | |
296 | * the page. | |
297 | * | |
298 | * General strategy: | |
299 | * Find all processes having the page mapped and kill them. | |
300 | * But we keep a page reference around so that the page is not | |
301 | * actually freed yet. | |
302 | * Then stash the page away | |
303 | * | |
304 | * There's no convenient way to get back to mapped processes | |
305 | * from the VMAs. So do a brute-force search over all | |
306 | * running processes. | |
307 | * | |
308 | * Remember that machine checks are not common (or rather | |
309 | * if they are common you have other problems), so this shouldn't | |
310 | * be a performance issue. | |
311 | * | |
312 | * Also there are some races possible while we get from the | |
313 | * error detection to actually handle it. | |
314 | */ | |
315 | ||
316 | struct to_kill { | |
317 | struct list_head nd; | |
318 | struct task_struct *tsk; | |
319 | unsigned long addr; | |
320 | short size_shift; | |
ae1139ec DW |
321 | }; |
322 | ||
6a46079c | 323 | /* |
7329bbeb TL |
324 | * Send all the processes who have the page mapped a signal. |
325 | * ``action optional'' if they are not immediately affected by the error | |
326 | * ``action required'' if error happened in current execution context | |
6a46079c | 327 | */ |
ae1139ec | 328 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 329 | { |
ae1139ec DW |
330 | struct task_struct *t = tk->tsk; |
331 | short addr_lsb = tk->size_shift; | |
872e9a20 | 332 | int ret = 0; |
6a46079c | 333 | |
96f96763 | 334 | pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", |
872e9a20 | 335 | pfn, t->comm, t->pid); |
7329bbeb | 336 | |
49775047 ML |
337 | if ((flags & MF_ACTION_REQUIRED) && (t == current)) |
338 | ret = force_sig_mceerr(BUS_MCEERR_AR, | |
339 | (void __user *)tk->addr, addr_lsb); | |
340 | else | |
7329bbeb | 341 | /* |
49775047 ML |
342 | * Signal other processes sharing the page if they have |
343 | * PF_MCE_EARLY set. | |
7329bbeb TL |
344 | * Don't use force here, it's convenient if the signal |
345 | * can be temporarily blocked. | |
346 | * This could cause a loop when the user sets SIGBUS | |
347 | * to SIG_IGN, but hopefully no one will do that? | |
348 | */ | |
ae1139ec | 349 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
9cf28191 | 350 | addr_lsb, t); |
6a46079c | 351 | if (ret < 0) |
96f96763 | 352 | pr_info("Error sending signal to %s:%d: %d\n", |
1170532b | 353 | t->comm, t->pid, ret); |
6a46079c AK |
354 | return ret; |
355 | } | |
356 | ||
588f9ce6 | 357 | /* |
47e431f4 | 358 | * Unknown page type encountered. Try to check whether it can turn PageLRU by |
d0505e9f | 359 | * lru_add_drain_all. |
588f9ce6 | 360 | */ |
d0505e9f | 361 | void shake_page(struct page *p) |
588f9ce6 | 362 | { |
8bcb74de NH |
363 | if (PageHuge(p)) |
364 | return; | |
588f9ce6 | 365 | /* |
d0505e9f YS |
366 | * TODO: Could shrink slab caches here if a lightweight range-based |
367 | * shrinker will be available. | |
588f9ce6 | 368 | */ |
b7b618da ML |
369 | if (PageSlab(p)) |
370 | return; | |
371 | ||
372 | lru_add_drain_all(); | |
588f9ce6 AK |
373 | } |
374 | EXPORT_SYMBOL_GPL(shake_page); | |
375 | ||
c36e2024 SR |
376 | static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct *vma, |
377 | unsigned long address) | |
6100e34b | 378 | { |
5c91c0e7 | 379 | unsigned long ret = 0; |
6100e34b DW |
380 | pgd_t *pgd; |
381 | p4d_t *p4d; | |
382 | pud_t *pud; | |
383 | pmd_t *pmd; | |
384 | pte_t *pte; | |
c33c7948 | 385 | pte_t ptent; |
6100e34b | 386 | |
a994402b | 387 | VM_BUG_ON_VMA(address == -EFAULT, vma); |
6100e34b DW |
388 | pgd = pgd_offset(vma->vm_mm, address); |
389 | if (!pgd_present(*pgd)) | |
390 | return 0; | |
391 | p4d = p4d_offset(pgd, address); | |
392 | if (!p4d_present(*p4d)) | |
393 | return 0; | |
394 | pud = pud_offset(p4d, address); | |
395 | if (!pud_present(*pud)) | |
396 | return 0; | |
397 | if (pud_devmap(*pud)) | |
398 | return PUD_SHIFT; | |
399 | pmd = pmd_offset(pud, address); | |
400 | if (!pmd_present(*pmd)) | |
401 | return 0; | |
402 | if (pmd_devmap(*pmd)) | |
403 | return PMD_SHIFT; | |
404 | pte = pte_offset_map(pmd, address); | |
04dee9e8 HD |
405 | if (!pte) |
406 | return 0; | |
c33c7948 RR |
407 | ptent = ptep_get(pte); |
408 | if (pte_present(ptent) && pte_devmap(ptent)) | |
5c91c0e7 QZ |
409 | ret = PAGE_SHIFT; |
410 | pte_unmap(pte); | |
411 | return ret; | |
6100e34b | 412 | } |
6a46079c AK |
413 | |
414 | /* | |
415 | * Failure handling: if we can't find or can't kill a process there's | |
416 | * not much we can do. We just print a message and ignore otherwise. | |
417 | */ | |
418 | ||
ac87ca0e DW |
419 | #define FSDAX_INVALID_PGOFF ULONG_MAX |
420 | ||
6a46079c AK |
421 | /* |
422 | * Schedule a process for later kill. | |
423 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
c36e2024 | 424 | * |
ac87ca0e DW |
425 | * Note: @fsdax_pgoff is used only when @p is a fsdax page and a |
426 | * filesystem with a memory failure handler has claimed the | |
427 | * memory_failure event. In all other cases, page->index and | |
428 | * page->mapping are sufficient for mapping the page back to its | |
429 | * corresponding user virtual address. | |
6a46079c | 430 | */ |
4f775086 LX |
431 | static void __add_to_kill(struct task_struct *tsk, struct page *p, |
432 | struct vm_area_struct *vma, struct list_head *to_kill, | |
433 | unsigned long ksm_addr, pgoff_t fsdax_pgoff) | |
6a46079c AK |
434 | { |
435 | struct to_kill *tk; | |
436 | ||
996ff7a0 JC |
437 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
438 | if (!tk) { | |
96f96763 | 439 | pr_err("Out of memory while machine check handling\n"); |
996ff7a0 | 440 | return; |
6a46079c | 441 | } |
996ff7a0 | 442 | |
4f775086 | 443 | tk->addr = ksm_addr ? ksm_addr : page_address_in_vma(p, vma); |
c36e2024 | 444 | if (is_zone_device_page(p)) { |
ac87ca0e | 445 | if (fsdax_pgoff != FSDAX_INVALID_PGOFF) |
c36e2024 SR |
446 | tk->addr = vma_pgoff_address(fsdax_pgoff, 1, vma); |
447 | tk->size_shift = dev_pagemap_mapping_shift(vma, tk->addr); | |
448 | } else | |
75068518 | 449 | tk->size_shift = page_shift(compound_head(p)); |
6a46079c AK |
450 | |
451 | /* | |
3d7fed4a JC |
452 | * Send SIGKILL if "tk->addr == -EFAULT". Also, as |
453 | * "tk->size_shift" is always non-zero for !is_zone_device_page(), | |
454 | * so "tk->size_shift == 0" effectively checks no mapping on | |
455 | * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times | |
456 | * to a process' address space, it's possible not all N VMAs | |
457 | * contain mappings for the page, but at least one VMA does. | |
458 | * Only deliver SIGBUS with payload derived from the VMA that | |
459 | * has a mapping for the page. | |
6a46079c | 460 | */ |
3d7fed4a | 461 | if (tk->addr == -EFAULT) { |
96f96763 | 462 | pr_info("Unable to find user space address %lx in %s\n", |
6a46079c | 463 | page_to_pfn(p), tsk->comm); |
3d7fed4a JC |
464 | } else if (tk->size_shift == 0) { |
465 | kfree(tk); | |
466 | return; | |
6a46079c | 467 | } |
996ff7a0 | 468 | |
6a46079c AK |
469 | get_task_struct(tsk); |
470 | tk->tsk = tsk; | |
471 | list_add_tail(&tk->nd, to_kill); | |
472 | } | |
473 | ||
4f775086 LX |
474 | static void add_to_kill_anon_file(struct task_struct *tsk, struct page *p, |
475 | struct vm_area_struct *vma, | |
476 | struct list_head *to_kill) | |
477 | { | |
478 | __add_to_kill(tsk, p, vma, to_kill, 0, FSDAX_INVALID_PGOFF); | |
479 | } | |
480 | ||
4248d008 LX |
481 | #ifdef CONFIG_KSM |
482 | static bool task_in_to_kill_list(struct list_head *to_kill, | |
483 | struct task_struct *tsk) | |
484 | { | |
485 | struct to_kill *tk, *next; | |
486 | ||
487 | list_for_each_entry_safe(tk, next, to_kill, nd) { | |
488 | if (tk->tsk == tsk) | |
489 | return true; | |
490 | } | |
491 | ||
492 | return false; | |
493 | } | |
494 | void add_to_kill_ksm(struct task_struct *tsk, struct page *p, | |
495 | struct vm_area_struct *vma, struct list_head *to_kill, | |
496 | unsigned long ksm_addr) | |
497 | { | |
498 | if (!task_in_to_kill_list(to_kill, tsk)) | |
499 | __add_to_kill(tsk, p, vma, to_kill, ksm_addr, FSDAX_INVALID_PGOFF); | |
500 | } | |
501 | #endif | |
6a46079c AK |
502 | /* |
503 | * Kill the processes that have been collected earlier. | |
504 | * | |
a21c184f ML |
505 | * Only do anything when FORCEKILL is set, otherwise just free the |
506 | * list (this is used for clean pages which do not need killing) | |
6a46079c AK |
507 | * Also when FAIL is set do a force kill because something went |
508 | * wrong earlier. | |
509 | */ | |
ae1139ec DW |
510 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
511 | unsigned long pfn, int flags) | |
6a46079c AK |
512 | { |
513 | struct to_kill *tk, *next; | |
514 | ||
54f9555d | 515 | list_for_each_entry_safe(tk, next, to_kill, nd) { |
6751ed65 | 516 | if (forcekill) { |
6a46079c | 517 | /* |
af901ca1 | 518 | * In case something went wrong with munmapping |
6a46079c AK |
519 | * make sure the process doesn't catch the |
520 | * signal and then access the memory. Just kill it. | |
6a46079c | 521 | */ |
3d7fed4a | 522 | if (fail || tk->addr == -EFAULT) { |
96f96763 | 523 | pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 524 | pfn, tk->tsk->comm, tk->tsk->pid); |
6376360e NH |
525 | do_send_sig_info(SIGKILL, SEND_SIG_PRIV, |
526 | tk->tsk, PIDTYPE_PID); | |
6a46079c AK |
527 | } |
528 | ||
529 | /* | |
530 | * In theory the process could have mapped | |
531 | * something else on the address in-between. We could | |
532 | * check for that, but we need to tell the | |
533 | * process anyways. | |
534 | */ | |
ae1139ec | 535 | else if (kill_proc(tk, pfn, flags) < 0) |
96f96763 | 536 | pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 537 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c | 538 | } |
54f9555d | 539 | list_del(&tk->nd); |
6a46079c AK |
540 | put_task_struct(tk->tsk); |
541 | kfree(tk); | |
542 | } | |
543 | } | |
544 | ||
3ba08129 NH |
545 | /* |
546 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
547 | * on behalf of the thread group. Return task_struct of the (first found) | |
548 | * dedicated thread if found, and return NULL otherwise. | |
549 | * | |
d256d1cd TT |
550 | * We already hold rcu lock in the caller, so we don't have to call |
551 | * rcu_read_lock/unlock() in this function. | |
3ba08129 NH |
552 | */ |
553 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 554 | { |
3ba08129 NH |
555 | struct task_struct *t; |
556 | ||
4e018b45 NH |
557 | for_each_thread(tsk, t) { |
558 | if (t->flags & PF_MCE_PROCESS) { | |
559 | if (t->flags & PF_MCE_EARLY) | |
560 | return t; | |
561 | } else { | |
562 | if (sysctl_memory_failure_early_kill) | |
563 | return t; | |
564 | } | |
565 | } | |
3ba08129 NH |
566 | return NULL; |
567 | } | |
568 | ||
569 | /* | |
570 | * Determine whether a given process is "early kill" process which expects | |
571 | * to be signaled when some page under the process is hwpoisoned. | |
572 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
30c9cf49 | 573 | * specified) if the process is "early kill" and otherwise returns NULL. |
03151c6e | 574 | * |
30c9cf49 AY |
575 | * Note that the above is true for Action Optional case. For Action Required |
576 | * case, it's only meaningful to the current thread which need to be signaled | |
577 | * with SIGBUS, this error is Action Optional for other non current | |
578 | * processes sharing the same error page,if the process is "early kill", the | |
579 | * task_struct of the dedicated thread will also be returned. | |
3ba08129 | 580 | */ |
4248d008 | 581 | struct task_struct *task_early_kill(struct task_struct *tsk, int force_early) |
3ba08129 | 582 | { |
6a46079c | 583 | if (!tsk->mm) |
3ba08129 | 584 | return NULL; |
30c9cf49 AY |
585 | /* |
586 | * Comparing ->mm here because current task might represent | |
587 | * a subthread, while tsk always points to the main thread. | |
588 | */ | |
589 | if (force_early && tsk->mm == current->mm) | |
590 | return current; | |
591 | ||
4e018b45 | 592 | return find_early_kill_thread(tsk); |
6a46079c AK |
593 | } |
594 | ||
595 | /* | |
596 | * Collect processes when the error hit an anonymous page. | |
597 | */ | |
598 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 599 | int force_early) |
6a46079c | 600 | { |
9595d769 | 601 | struct folio *folio = page_folio(page); |
6a46079c AK |
602 | struct vm_area_struct *vma; |
603 | struct task_struct *tsk; | |
604 | struct anon_vma *av; | |
bf181b9f | 605 | pgoff_t pgoff; |
6a46079c | 606 | |
6d4675e6 | 607 | av = folio_lock_anon_vma_read(folio, NULL); |
6a46079c | 608 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
609 | return; |
610 | ||
a0f7a756 | 611 | pgoff = page_to_pgoff(page); |
d256d1cd | 612 | rcu_read_lock(); |
5885c6a6 | 613 | for_each_process(tsk) { |
5beb4930 | 614 | struct anon_vma_chain *vmac; |
3ba08129 | 615 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 616 | |
3ba08129 | 617 | if (!t) |
6a46079c | 618 | continue; |
bf181b9f ML |
619 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
620 | pgoff, pgoff) { | |
5beb4930 | 621 | vma = vmac->vma; |
36537a67 ML |
622 | if (vma->vm_mm != t->mm) |
623 | continue; | |
6a46079c AK |
624 | if (!page_mapped_in_vma(page, vma)) |
625 | continue; | |
4f775086 | 626 | add_to_kill_anon_file(t, page, vma, to_kill); |
6a46079c AK |
627 | } |
628 | } | |
d256d1cd | 629 | rcu_read_unlock(); |
0c826c0b | 630 | anon_vma_unlock_read(av); |
6a46079c AK |
631 | } |
632 | ||
633 | /* | |
634 | * Collect processes when the error hit a file mapped page. | |
635 | */ | |
636 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 637 | int force_early) |
6a46079c AK |
638 | { |
639 | struct vm_area_struct *vma; | |
640 | struct task_struct *tsk; | |
6a46079c | 641 | struct address_space *mapping = page->mapping; |
c43bc03d | 642 | pgoff_t pgoff; |
6a46079c | 643 | |
d28eb9c8 | 644 | i_mmap_lock_read(mapping); |
d256d1cd | 645 | rcu_read_lock(); |
c43bc03d | 646 | pgoff = page_to_pgoff(page); |
6a46079c | 647 | for_each_process(tsk) { |
3ba08129 | 648 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 649 | |
3ba08129 | 650 | if (!t) |
6a46079c | 651 | continue; |
6b2dbba8 | 652 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
653 | pgoff) { |
654 | /* | |
655 | * Send early kill signal to tasks where a vma covers | |
656 | * the page but the corrupted page is not necessarily | |
5885c6a6 | 657 | * mapped in its pte. |
6a46079c AK |
658 | * Assume applications who requested early kill want |
659 | * to be informed of all such data corruptions. | |
660 | */ | |
3ba08129 | 661 | if (vma->vm_mm == t->mm) |
4f775086 | 662 | add_to_kill_anon_file(t, page, vma, to_kill); |
6a46079c AK |
663 | } |
664 | } | |
d256d1cd | 665 | rcu_read_unlock(); |
d28eb9c8 | 666 | i_mmap_unlock_read(mapping); |
6a46079c AK |
667 | } |
668 | ||
c36e2024 | 669 | #ifdef CONFIG_FS_DAX |
4f775086 LX |
670 | static void add_to_kill_fsdax(struct task_struct *tsk, struct page *p, |
671 | struct vm_area_struct *vma, | |
672 | struct list_head *to_kill, pgoff_t pgoff) | |
673 | { | |
674 | __add_to_kill(tsk, p, vma, to_kill, 0, pgoff); | |
675 | } | |
676 | ||
c36e2024 SR |
677 | /* |
678 | * Collect processes when the error hit a fsdax page. | |
679 | */ | |
680 | static void collect_procs_fsdax(struct page *page, | |
681 | struct address_space *mapping, pgoff_t pgoff, | |
682 | struct list_head *to_kill) | |
683 | { | |
684 | struct vm_area_struct *vma; | |
685 | struct task_struct *tsk; | |
686 | ||
687 | i_mmap_lock_read(mapping); | |
d256d1cd | 688 | rcu_read_lock(); |
c36e2024 SR |
689 | for_each_process(tsk) { |
690 | struct task_struct *t = task_early_kill(tsk, true); | |
691 | ||
692 | if (!t) | |
693 | continue; | |
694 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { | |
695 | if (vma->vm_mm == t->mm) | |
4f775086 | 696 | add_to_kill_fsdax(t, page, vma, to_kill, pgoff); |
c36e2024 SR |
697 | } |
698 | } | |
d256d1cd | 699 | rcu_read_unlock(); |
c36e2024 SR |
700 | i_mmap_unlock_read(mapping); |
701 | } | |
702 | #endif /* CONFIG_FS_DAX */ | |
703 | ||
6a46079c AK |
704 | /* |
705 | * Collect the processes who have the corrupted page mapped to kill. | |
6a46079c | 706 | */ |
74614de1 TL |
707 | static void collect_procs(struct page *page, struct list_head *tokill, |
708 | int force_early) | |
6a46079c | 709 | { |
6a46079c AK |
710 | if (!page->mapping) |
711 | return; | |
4248d008 LX |
712 | if (unlikely(PageKsm(page))) |
713 | collect_procs_ksm(page, tokill, force_early); | |
714 | else if (PageAnon(page)) | |
996ff7a0 | 715 | collect_procs_anon(page, tokill, force_early); |
6a46079c | 716 | else |
996ff7a0 | 717 | collect_procs_file(page, tokill, force_early); |
6a46079c AK |
718 | } |
719 | ||
6885938c | 720 | struct hwpoison_walk { |
a3f5d80e NH |
721 | struct to_kill tk; |
722 | unsigned long pfn; | |
723 | int flags; | |
724 | }; | |
725 | ||
726 | static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift) | |
727 | { | |
728 | tk->addr = addr; | |
729 | tk->size_shift = shift; | |
730 | } | |
731 | ||
732 | static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift, | |
733 | unsigned long poisoned_pfn, struct to_kill *tk) | |
734 | { | |
735 | unsigned long pfn = 0; | |
736 | ||
737 | if (pte_present(pte)) { | |
738 | pfn = pte_pfn(pte); | |
739 | } else { | |
740 | swp_entry_t swp = pte_to_swp_entry(pte); | |
741 | ||
742 | if (is_hwpoison_entry(swp)) | |
0d206b5d | 743 | pfn = swp_offset_pfn(swp); |
a3f5d80e NH |
744 | } |
745 | ||
746 | if (!pfn || pfn != poisoned_pfn) | |
747 | return 0; | |
748 | ||
749 | set_to_kill(tk, addr, shift); | |
750 | return 1; | |
751 | } | |
752 | ||
753 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
754 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
6885938c | 755 | struct hwpoison_walk *hwp) |
a3f5d80e NH |
756 | { |
757 | pmd_t pmd = *pmdp; | |
758 | unsigned long pfn; | |
759 | unsigned long hwpoison_vaddr; | |
760 | ||
761 | if (!pmd_present(pmd)) | |
762 | return 0; | |
763 | pfn = pmd_pfn(pmd); | |
764 | if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) { | |
765 | hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT); | |
766 | set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT); | |
767 | return 1; | |
768 | } | |
769 | return 0; | |
770 | } | |
771 | #else | |
772 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
6885938c | 773 | struct hwpoison_walk *hwp) |
a3f5d80e NH |
774 | { |
775 | return 0; | |
776 | } | |
777 | #endif | |
778 | ||
779 | static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr, | |
780 | unsigned long end, struct mm_walk *walk) | |
781 | { | |
6885938c | 782 | struct hwpoison_walk *hwp = walk->private; |
a3f5d80e | 783 | int ret = 0; |
ea3732f7 | 784 | pte_t *ptep, *mapped_pte; |
a3f5d80e NH |
785 | spinlock_t *ptl; |
786 | ||
787 | ptl = pmd_trans_huge_lock(pmdp, walk->vma); | |
788 | if (ptl) { | |
789 | ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp); | |
790 | spin_unlock(ptl); | |
791 | goto out; | |
792 | } | |
793 | ||
ea3732f7 ML |
794 | mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, |
795 | addr, &ptl); | |
04dee9e8 HD |
796 | if (!ptep) |
797 | goto out; | |
798 | ||
a3f5d80e | 799 | for (; addr != end; ptep++, addr += PAGE_SIZE) { |
c33c7948 | 800 | ret = check_hwpoisoned_entry(ptep_get(ptep), addr, PAGE_SHIFT, |
a3f5d80e NH |
801 | hwp->pfn, &hwp->tk); |
802 | if (ret == 1) | |
803 | break; | |
804 | } | |
ea3732f7 | 805 | pte_unmap_unlock(mapped_pte, ptl); |
a3f5d80e NH |
806 | out: |
807 | cond_resched(); | |
808 | return ret; | |
809 | } | |
810 | ||
811 | #ifdef CONFIG_HUGETLB_PAGE | |
812 | static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask, | |
813 | unsigned long addr, unsigned long end, | |
814 | struct mm_walk *walk) | |
815 | { | |
6885938c | 816 | struct hwpoison_walk *hwp = walk->private; |
a3f5d80e NH |
817 | pte_t pte = huge_ptep_get(ptep); |
818 | struct hstate *h = hstate_vma(walk->vma); | |
819 | ||
820 | return check_hwpoisoned_entry(pte, addr, huge_page_shift(h), | |
821 | hwp->pfn, &hwp->tk); | |
822 | } | |
823 | #else | |
824 | #define hwpoison_hugetlb_range NULL | |
825 | #endif | |
826 | ||
6885938c | 827 | static const struct mm_walk_ops hwpoison_walk_ops = { |
a3f5d80e NH |
828 | .pmd_entry = hwpoison_pte_range, |
829 | .hugetlb_entry = hwpoison_hugetlb_range, | |
49b06385 | 830 | .walk_lock = PGWALK_RDLOCK, |
a3f5d80e NH |
831 | }; |
832 | ||
833 | /* | |
834 | * Sends SIGBUS to the current process with error info. | |
835 | * | |
836 | * This function is intended to handle "Action Required" MCEs on already | |
837 | * hardware poisoned pages. They could happen, for example, when | |
838 | * memory_failure() failed to unmap the error page at the first call, or | |
839 | * when multiple local machine checks happened on different CPUs. | |
840 | * | |
841 | * MCE handler currently has no easy access to the error virtual address, | |
842 | * so this function walks page table to find it. The returned virtual address | |
843 | * is proper in most cases, but it could be wrong when the application | |
844 | * process has multiple entries mapping the error page. | |
845 | */ | |
846 | static int kill_accessing_process(struct task_struct *p, unsigned long pfn, | |
847 | int flags) | |
848 | { | |
849 | int ret; | |
6885938c | 850 | struct hwpoison_walk priv = { |
a3f5d80e NH |
851 | .pfn = pfn, |
852 | }; | |
853 | priv.tk.tsk = p; | |
854 | ||
77677cdb SX |
855 | if (!p->mm) |
856 | return -EFAULT; | |
857 | ||
a3f5d80e | 858 | mmap_read_lock(p->mm); |
6885938c | 859 | ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwpoison_walk_ops, |
a3f5d80e NH |
860 | (void *)&priv); |
861 | if (ret == 1 && priv.tk.addr) | |
862 | kill_proc(&priv.tk, pfn, flags); | |
046545a6 NH |
863 | else |
864 | ret = 0; | |
a3f5d80e | 865 | mmap_read_unlock(p->mm); |
046545a6 | 866 | return ret > 0 ? -EHWPOISON : -EFAULT; |
a3f5d80e NH |
867 | } |
868 | ||
6a46079c | 869 | static const char *action_name[] = { |
cc637b17 XX |
870 | [MF_IGNORED] = "Ignored", |
871 | [MF_FAILED] = "Failed", | |
872 | [MF_DELAYED] = "Delayed", | |
873 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
874 | }; |
875 | ||
876 | static const char * const action_page_types[] = { | |
cc637b17 XX |
877 | [MF_MSG_KERNEL] = "reserved kernel page", |
878 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
879 | [MF_MSG_SLAB] = "kernel slab page", | |
880 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
cc637b17 XX |
881 | [MF_MSG_HUGE] = "huge page", |
882 | [MF_MSG_FREE_HUGE] = "free huge page", | |
883 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", | |
884 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
885 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
886 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
887 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
888 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
889 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
890 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
891 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
892 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
893 | [MF_MSG_BUDDY] = "free buddy page", | |
6100e34b | 894 | [MF_MSG_DAX] = "dax page", |
5d1fd5dc | 895 | [MF_MSG_UNSPLIT_THP] = "unsplit thp", |
cc637b17 | 896 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
897 | }; |
898 | ||
dc2a1cbf WF |
899 | /* |
900 | * XXX: It is possible that a page is isolated from LRU cache, | |
901 | * and then kept in swap cache or failed to remove from page cache. | |
902 | * The page count will stop it from being freed by unpoison. | |
903 | * Stress tests should be aware of this memory leak problem. | |
904 | */ | |
905 | static int delete_from_lru_cache(struct page *p) | |
906 | { | |
f7f9c00d | 907 | if (isolate_lru_page(p)) { |
dc2a1cbf WF |
908 | /* |
909 | * Clear sensible page flags, so that the buddy system won't | |
910 | * complain when the page is unpoison-and-freed. | |
911 | */ | |
912 | ClearPageActive(p); | |
913 | ClearPageUnevictable(p); | |
18365225 MH |
914 | |
915 | /* | |
916 | * Poisoned page might never drop its ref count to 0 so we have | |
917 | * to uncharge it manually from its memcg. | |
918 | */ | |
bbc6b703 | 919 | mem_cgroup_uncharge(page_folio(p)); |
18365225 | 920 | |
dc2a1cbf WF |
921 | /* |
922 | * drop the page count elevated by isolate_lru_page() | |
923 | */ | |
09cbfeaf | 924 | put_page(p); |
dc2a1cbf WF |
925 | return 0; |
926 | } | |
927 | return -EIO; | |
928 | } | |
929 | ||
78bb9203 NH |
930 | static int truncate_error_page(struct page *p, unsigned long pfn, |
931 | struct address_space *mapping) | |
932 | { | |
933 | int ret = MF_FAILED; | |
934 | ||
935 | if (mapping->a_ops->error_remove_page) { | |
ac5efa78 | 936 | struct folio *folio = page_folio(p); |
78bb9203 NH |
937 | int err = mapping->a_ops->error_remove_page(mapping, p); |
938 | ||
0201ebf2 | 939 | if (err != 0) |
96f96763 | 940 | pr_info("%#lx: Failed to punch page: %d\n", pfn, err); |
0201ebf2 | 941 | else if (!filemap_release_folio(folio, GFP_NOIO)) |
96f96763 | 942 | pr_info("%#lx: failed to release buffers\n", pfn); |
0201ebf2 | 943 | else |
78bb9203 | 944 | ret = MF_RECOVERED; |
78bb9203 NH |
945 | } else { |
946 | /* | |
947 | * If the file system doesn't support it just invalidate | |
948 | * This fails on dirty or anything with private pages | |
949 | */ | |
950 | if (invalidate_inode_page(p)) | |
951 | ret = MF_RECOVERED; | |
952 | else | |
96f96763 | 953 | pr_info("%#lx: Failed to invalidate\n", pfn); |
78bb9203 NH |
954 | } |
955 | ||
956 | return ret; | |
957 | } | |
958 | ||
dd0f230a YS |
959 | struct page_state { |
960 | unsigned long mask; | |
961 | unsigned long res; | |
962 | enum mf_action_page_type type; | |
963 | ||
964 | /* Callback ->action() has to unlock the relevant page inside it. */ | |
965 | int (*action)(struct page_state *ps, struct page *p); | |
966 | }; | |
967 | ||
968 | /* | |
969 | * Return true if page is still referenced by others, otherwise return | |
970 | * false. | |
971 | * | |
972 | * The extra_pins is true when one extra refcount is expected. | |
973 | */ | |
974 | static bool has_extra_refcount(struct page_state *ps, struct page *p, | |
975 | bool extra_pins) | |
976 | { | |
977 | int count = page_count(p) - 1; | |
978 | ||
979 | if (extra_pins) | |
980 | count -= 1; | |
981 | ||
982 | if (count > 0) { | |
96f96763 | 983 | pr_err("%#lx: %s still referenced by %d users\n", |
dd0f230a YS |
984 | page_to_pfn(p), action_page_types[ps->type], count); |
985 | return true; | |
986 | } | |
987 | ||
988 | return false; | |
989 | } | |
990 | ||
6a46079c AK |
991 | /* |
992 | * Error hit kernel page. | |
993 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
994 | * could be more sophisticated. | |
995 | */ | |
dd0f230a | 996 | static int me_kernel(struct page_state *ps, struct page *p) |
6a46079c | 997 | { |
ea6d0630 | 998 | unlock_page(p); |
cc637b17 | 999 | return MF_IGNORED; |
6a46079c AK |
1000 | } |
1001 | ||
1002 | /* | |
1003 | * Page in unknown state. Do nothing. | |
1004 | */ | |
dd0f230a | 1005 | static int me_unknown(struct page_state *ps, struct page *p) |
6a46079c | 1006 | { |
96f96763 | 1007 | pr_err("%#lx: Unknown page state\n", page_to_pfn(p)); |
ea6d0630 | 1008 | unlock_page(p); |
cc637b17 | 1009 | return MF_FAILED; |
6a46079c AK |
1010 | } |
1011 | ||
6a46079c AK |
1012 | /* |
1013 | * Clean (or cleaned) page cache page. | |
1014 | */ | |
dd0f230a | 1015 | static int me_pagecache_clean(struct page_state *ps, struct page *p) |
6a46079c | 1016 | { |
ea6d0630 | 1017 | int ret; |
6a46079c | 1018 | struct address_space *mapping; |
a7605426 | 1019 | bool extra_pins; |
6a46079c | 1020 | |
dc2a1cbf WF |
1021 | delete_from_lru_cache(p); |
1022 | ||
6a46079c AK |
1023 | /* |
1024 | * For anonymous pages we're done the only reference left | |
1025 | * should be the one m_f() holds. | |
1026 | */ | |
ea6d0630 NH |
1027 | if (PageAnon(p)) { |
1028 | ret = MF_RECOVERED; | |
1029 | goto out; | |
1030 | } | |
6a46079c AK |
1031 | |
1032 | /* | |
1033 | * Now truncate the page in the page cache. This is really | |
1034 | * more like a "temporary hole punch" | |
1035 | * Don't do this for block devices when someone else | |
1036 | * has a reference, because it could be file system metadata | |
1037 | * and that's not safe to truncate. | |
1038 | */ | |
1039 | mapping = page_mapping(p); | |
1040 | if (!mapping) { | |
1041 | /* | |
1042 | * Page has been teared down in the meanwhile | |
1043 | */ | |
ea6d0630 NH |
1044 | ret = MF_FAILED; |
1045 | goto out; | |
6a46079c AK |
1046 | } |
1047 | ||
a7605426 YS |
1048 | /* |
1049 | * The shmem page is kept in page cache instead of truncating | |
1050 | * so is expected to have an extra refcount after error-handling. | |
1051 | */ | |
1052 | extra_pins = shmem_mapping(mapping); | |
1053 | ||
6a46079c AK |
1054 | /* |
1055 | * Truncation is a bit tricky. Enable it per file system for now. | |
1056 | * | |
9608703e | 1057 | * Open: to take i_rwsem or not for this? Right now we don't. |
6a46079c | 1058 | */ |
dd0f230a | 1059 | ret = truncate_error_page(p, page_to_pfn(p), mapping); |
a7605426 YS |
1060 | if (has_extra_refcount(ps, p, extra_pins)) |
1061 | ret = MF_FAILED; | |
1062 | ||
ea6d0630 NH |
1063 | out: |
1064 | unlock_page(p); | |
dd0f230a | 1065 | |
ea6d0630 | 1066 | return ret; |
6a46079c AK |
1067 | } |
1068 | ||
1069 | /* | |
549543df | 1070 | * Dirty pagecache page |
6a46079c AK |
1071 | * Issues: when the error hit a hole page the error is not properly |
1072 | * propagated. | |
1073 | */ | |
dd0f230a | 1074 | static int me_pagecache_dirty(struct page_state *ps, struct page *p) |
6a46079c AK |
1075 | { |
1076 | struct address_space *mapping = page_mapping(p); | |
1077 | ||
1078 | SetPageError(p); | |
1079 | /* TBD: print more information about the file. */ | |
1080 | if (mapping) { | |
1081 | /* | |
1082 | * IO error will be reported by write(), fsync(), etc. | |
1083 | * who check the mapping. | |
1084 | * This way the application knows that something went | |
1085 | * wrong with its dirty file data. | |
1086 | * | |
1087 | * There's one open issue: | |
1088 | * | |
1089 | * The EIO will be only reported on the next IO | |
1090 | * operation and then cleared through the IO map. | |
1091 | * Normally Linux has two mechanisms to pass IO error | |
1092 | * first through the AS_EIO flag in the address space | |
1093 | * and then through the PageError flag in the page. | |
1094 | * Since we drop pages on memory failure handling the | |
1095 | * only mechanism open to use is through AS_AIO. | |
1096 | * | |
1097 | * This has the disadvantage that it gets cleared on | |
1098 | * the first operation that returns an error, while | |
1099 | * the PageError bit is more sticky and only cleared | |
1100 | * when the page is reread or dropped. If an | |
1101 | * application assumes it will always get error on | |
1102 | * fsync, but does other operations on the fd before | |
25985edc | 1103 | * and the page is dropped between then the error |
6a46079c AK |
1104 | * will not be properly reported. |
1105 | * | |
1106 | * This can already happen even without hwpoisoned | |
1107 | * pages: first on metadata IO errors (which only | |
1108 | * report through AS_EIO) or when the page is dropped | |
1109 | * at the wrong time. | |
1110 | * | |
1111 | * So right now we assume that the application DTRT on | |
1112 | * the first EIO, but we're not worse than other parts | |
1113 | * of the kernel. | |
1114 | */ | |
af21bfaf | 1115 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
1116 | } |
1117 | ||
dd0f230a | 1118 | return me_pagecache_clean(ps, p); |
6a46079c AK |
1119 | } |
1120 | ||
1121 | /* | |
1122 | * Clean and dirty swap cache. | |
1123 | * | |
1124 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
1125 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
1126 | * referenced concurrently by 2 types of PTEs: | |
1127 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
6da6b1d4 | 1128 | * try_to_unmap(!TTU_HWPOISON) to convert the normal PTEs to swap PTEs, |
6a46079c AK |
1129 | * and then |
1130 | * - clear dirty bit to prevent IO | |
1131 | * - remove from LRU | |
1132 | * - but keep in the swap cache, so that when we return to it on | |
1133 | * a later page fault, we know the application is accessing | |
1134 | * corrupted data and shall be killed (we installed simple | |
1135 | * interception code in do_swap_page to catch it). | |
1136 | * | |
1137 | * Clean swap cache pages can be directly isolated. A later page fault will | |
1138 | * bring in the known good data from disk. | |
1139 | */ | |
dd0f230a | 1140 | static int me_swapcache_dirty(struct page_state *ps, struct page *p) |
6a46079c | 1141 | { |
ea6d0630 | 1142 | int ret; |
dd0f230a | 1143 | bool extra_pins = false; |
ea6d0630 | 1144 | |
6a46079c AK |
1145 | ClearPageDirty(p); |
1146 | /* Trigger EIO in shmem: */ | |
1147 | ClearPageUptodate(p); | |
1148 | ||
ea6d0630 NH |
1149 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED; |
1150 | unlock_page(p); | |
dd0f230a YS |
1151 | |
1152 | if (ret == MF_DELAYED) | |
1153 | extra_pins = true; | |
1154 | ||
1155 | if (has_extra_refcount(ps, p, extra_pins)) | |
1156 | ret = MF_FAILED; | |
1157 | ||
ea6d0630 | 1158 | return ret; |
6a46079c AK |
1159 | } |
1160 | ||
dd0f230a | 1161 | static int me_swapcache_clean(struct page_state *ps, struct page *p) |
6a46079c | 1162 | { |
75fa68a5 | 1163 | struct folio *folio = page_folio(p); |
ea6d0630 NH |
1164 | int ret; |
1165 | ||
75fa68a5 | 1166 | delete_from_swap_cache(folio); |
e43c3afb | 1167 | |
ea6d0630 | 1168 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED; |
75fa68a5 | 1169 | folio_unlock(folio); |
dd0f230a YS |
1170 | |
1171 | if (has_extra_refcount(ps, p, false)) | |
1172 | ret = MF_FAILED; | |
1173 | ||
ea6d0630 | 1174 | return ret; |
6a46079c AK |
1175 | } |
1176 | ||
1177 | /* | |
1178 | * Huge pages. Needs work. | |
1179 | * Issues: | |
93f70f90 NH |
1180 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
1181 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c | 1182 | */ |
dd0f230a | 1183 | static int me_huge_page(struct page_state *ps, struct page *p) |
6a46079c | 1184 | { |
a8b2c2ce | 1185 | int res; |
93f70f90 | 1186 | struct page *hpage = compound_head(p); |
78bb9203 | 1187 | struct address_space *mapping; |
8625147c | 1188 | bool extra_pins = false; |
2491ffee | 1189 | |
78bb9203 NH |
1190 | mapping = page_mapping(hpage); |
1191 | if (mapping) { | |
dd0f230a | 1192 | res = truncate_error_page(hpage, page_to_pfn(p), mapping); |
8625147c JH |
1193 | /* The page is kept in page cache. */ |
1194 | extra_pins = true; | |
ea6d0630 | 1195 | unlock_page(hpage); |
78bb9203 NH |
1196 | } else { |
1197 | unlock_page(hpage); | |
1198 | /* | |
ef526b17 ML |
1199 | * migration entry prevents later access on error hugepage, |
1200 | * so we can free and dissolve it into buddy to save healthy | |
1201 | * subpages. | |
78bb9203 | 1202 | */ |
ef526b17 | 1203 | put_page(hpage); |
ceaf8fbe | 1204 | if (__page_handle_poison(p) >= 0) { |
a8b2c2ce OS |
1205 | page_ref_inc(p); |
1206 | res = MF_RECOVERED; | |
ceaf8fbe NH |
1207 | } else { |
1208 | res = MF_FAILED; | |
a8b2c2ce | 1209 | } |
93f70f90 | 1210 | } |
78bb9203 | 1211 | |
8625147c | 1212 | if (has_extra_refcount(ps, p, extra_pins)) |
dd0f230a YS |
1213 | res = MF_FAILED; |
1214 | ||
78bb9203 | 1215 | return res; |
6a46079c AK |
1216 | } |
1217 | ||
1218 | /* | |
1219 | * Various page states we can handle. | |
1220 | * | |
1221 | * A page state is defined by its current page->flags bits. | |
1222 | * The table matches them in order and calls the right handler. | |
1223 | * | |
1224 | * This is quite tricky because we can access page at any time | |
25985edc | 1225 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
1226 | * |
1227 | * This is not complete. More states could be added. | |
1228 | * For any missing state don't attempt recovery. | |
1229 | */ | |
1230 | ||
1231 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 1232 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
1233 | #define unevict (1UL << PG_unevictable) |
1234 | #define mlock (1UL << PG_mlocked) | |
6a46079c | 1235 | #define lru (1UL << PG_lru) |
6a46079c | 1236 | #define head (1UL << PG_head) |
6a46079c | 1237 | #define slab (1UL << PG_slab) |
6a46079c AK |
1238 | #define reserved (1UL << PG_reserved) |
1239 | ||
dd0f230a | 1240 | static struct page_state error_states[] = { |
cc637b17 | 1241 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
1242 | /* |
1243 | * free pages are specially detected outside this table: | |
1244 | * PG_buddy pages only make a small fraction of all free pages. | |
1245 | */ | |
6a46079c AK |
1246 | |
1247 | /* | |
1248 | * Could in theory check if slab page is free or if we can drop | |
1249 | * currently unused objects without touching them. But just | |
1250 | * treat it as standard kernel for now. | |
1251 | */ | |
cc637b17 | 1252 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 1253 | |
cc637b17 | 1254 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 1255 | |
cc637b17 XX |
1256 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
1257 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 1258 | |
cc637b17 XX |
1259 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
1260 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 1261 | |
cc637b17 XX |
1262 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
1263 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 1264 | |
cc637b17 XX |
1265 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
1266 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
1267 | |
1268 | /* | |
1269 | * Catchall entry: must be at end. | |
1270 | */ | |
cc637b17 | 1271 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
1272 | }; |
1273 | ||
2326c467 AK |
1274 | #undef dirty |
1275 | #undef sc | |
1276 | #undef unevict | |
1277 | #undef mlock | |
2326c467 | 1278 | #undef lru |
2326c467 | 1279 | #undef head |
2326c467 AK |
1280 | #undef slab |
1281 | #undef reserved | |
1282 | ||
18f41fa6 JY |
1283 | static void update_per_node_mf_stats(unsigned long pfn, |
1284 | enum mf_result result) | |
1285 | { | |
1286 | int nid = MAX_NUMNODES; | |
1287 | struct memory_failure_stats *mf_stats = NULL; | |
1288 | ||
1289 | nid = pfn_to_nid(pfn); | |
1290 | if (unlikely(nid < 0 || nid >= MAX_NUMNODES)) { | |
1291 | WARN_ONCE(1, "Memory failure: pfn=%#lx, invalid nid=%d", pfn, nid); | |
1292 | return; | |
1293 | } | |
1294 | ||
1295 | mf_stats = &NODE_DATA(nid)->mf_stats; | |
1296 | switch (result) { | |
1297 | case MF_IGNORED: | |
1298 | ++mf_stats->ignored; | |
1299 | break; | |
1300 | case MF_FAILED: | |
1301 | ++mf_stats->failed; | |
1302 | break; | |
1303 | case MF_DELAYED: | |
1304 | ++mf_stats->delayed; | |
1305 | break; | |
1306 | case MF_RECOVERED: | |
1307 | ++mf_stats->recovered; | |
1308 | break; | |
1309 | default: | |
1310 | WARN_ONCE(1, "Memory failure: mf_result=%d is not properly handled", result); | |
1311 | break; | |
1312 | } | |
1313 | ++mf_stats->total; | |
1314 | } | |
1315 | ||
ff604cf6 NH |
1316 | /* |
1317 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
1318 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
1319 | */ | |
b66d00df KW |
1320 | static int action_result(unsigned long pfn, enum mf_action_page_type type, |
1321 | enum mf_result result) | |
6a46079c | 1322 | { |
97f0b134 XX |
1323 | trace_memory_failure_event(pfn, type, result); |
1324 | ||
a46c9304 | 1325 | num_poisoned_pages_inc(pfn); |
18f41fa6 JY |
1326 | |
1327 | update_per_node_mf_stats(pfn, result); | |
1328 | ||
96f96763 | 1329 | pr_err("%#lx: recovery action for %s: %s\n", |
64d37a2b | 1330 | pfn, action_page_types[type], action_name[result]); |
b66d00df KW |
1331 | |
1332 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; | |
6a46079c AK |
1333 | } |
1334 | ||
1335 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 1336 | unsigned long pfn) |
6a46079c AK |
1337 | { |
1338 | int result; | |
1339 | ||
ea6d0630 | 1340 | /* page p should be unlocked after returning from ps->action(). */ |
dd0f230a | 1341 | result = ps->action(ps, p); |
7456b040 | 1342 | |
6a46079c AK |
1343 | /* Could do more checks here if page looks ok */ |
1344 | /* | |
1345 | * Could adjust zone counters here to correct for the missing page. | |
1346 | */ | |
1347 | ||
b66d00df | 1348 | return action_result(pfn, ps->type, result); |
6a46079c AK |
1349 | } |
1350 | ||
bf181c58 NH |
1351 | static inline bool PageHWPoisonTakenOff(struct page *page) |
1352 | { | |
1353 | return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON; | |
1354 | } | |
1355 | ||
1356 | void SetPageHWPoisonTakenOff(struct page *page) | |
1357 | { | |
1358 | set_page_private(page, MAGIC_HWPOISON); | |
1359 | } | |
1360 | ||
1361 | void ClearPageHWPoisonTakenOff(struct page *page) | |
1362 | { | |
1363 | if (PageHWPoison(page)) | |
1364 | set_page_private(page, 0); | |
1365 | } | |
1366 | ||
25182f05 NH |
1367 | /* |
1368 | * Return true if a page type of a given page is supported by hwpoison | |
1369 | * mechanism (while handling could fail), otherwise false. This function | |
1370 | * does not return true for hugetlb or device memory pages, so it's assumed | |
1371 | * to be called only in the context where we never have such pages. | |
1372 | */ | |
bf6445bc | 1373 | static inline bool HWPoisonHandlable(struct page *page, unsigned long flags) |
25182f05 | 1374 | { |
3f871370 | 1375 | /* Soft offline could migrate non-LRU movable pages */ |
bf6445bc | 1376 | if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page)) |
3f871370 | 1377 | return true; |
bf6445bc | 1378 | |
3f871370 | 1379 | return PageLRU(page) || is_free_buddy_page(page); |
25182f05 NH |
1380 | } |
1381 | ||
bf6445bc | 1382 | static int __get_hwpoison_page(struct page *page, unsigned long flags) |
ead07f6a | 1383 | { |
04bac040 | 1384 | struct folio *folio = page_folio(page); |
25182f05 NH |
1385 | int ret = 0; |
1386 | bool hugetlb = false; | |
1387 | ||
04bac040 | 1388 | ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, false); |
d31155b8 ML |
1389 | if (hugetlb) { |
1390 | /* Make sure hugetlb demotion did not happen from under us. */ | |
1391 | if (folio == page_folio(page)) | |
1392 | return ret; | |
1393 | if (ret > 0) { | |
1394 | folio_put(folio); | |
1395 | folio = page_folio(page); | |
1396 | } | |
1397 | } | |
25182f05 NH |
1398 | |
1399 | /* | |
04bac040 SK |
1400 | * This check prevents from calling folio_try_get() for any |
1401 | * unsupported type of folio in order to reduce the risk of unexpected | |
1402 | * races caused by taking a folio refcount. | |
25182f05 | 1403 | */ |
04bac040 | 1404 | if (!HWPoisonHandlable(&folio->page, flags)) |
fcc00621 | 1405 | return -EBUSY; |
ead07f6a | 1406 | |
04bac040 SK |
1407 | if (folio_try_get(folio)) { |
1408 | if (folio == page_folio(page)) | |
c2e7e00b KK |
1409 | return 1; |
1410 | ||
96f96763 | 1411 | pr_info("%#lx cannot catch tail\n", page_to_pfn(page)); |
04bac040 | 1412 | folio_put(folio); |
c2e7e00b KK |
1413 | } |
1414 | ||
1415 | return 0; | |
ead07f6a | 1416 | } |
ead07f6a | 1417 | |
2f714160 | 1418 | static int get_any_page(struct page *p, unsigned long flags) |
17e395b6 | 1419 | { |
2f714160 OS |
1420 | int ret = 0, pass = 0; |
1421 | bool count_increased = false; | |
17e395b6 | 1422 | |
2f714160 OS |
1423 | if (flags & MF_COUNT_INCREASED) |
1424 | count_increased = true; | |
1425 | ||
1426 | try_again: | |
0ed950d1 | 1427 | if (!count_increased) { |
bf6445bc | 1428 | ret = __get_hwpoison_page(p, flags); |
0ed950d1 NH |
1429 | if (!ret) { |
1430 | if (page_count(p)) { | |
1431 | /* We raced with an allocation, retry. */ | |
1432 | if (pass++ < 3) | |
1433 | goto try_again; | |
1434 | ret = -EBUSY; | |
1435 | } else if (!PageHuge(p) && !is_free_buddy_page(p)) { | |
1436 | /* We raced with put_page, retry. */ | |
1437 | if (pass++ < 3) | |
1438 | goto try_again; | |
1439 | ret = -EIO; | |
1440 | } | |
1441 | goto out; | |
1442 | } else if (ret == -EBUSY) { | |
fcc00621 NH |
1443 | /* |
1444 | * We raced with (possibly temporary) unhandlable | |
1445 | * page, retry. | |
1446 | */ | |
1447 | if (pass++ < 3) { | |
d0505e9f | 1448 | shake_page(p); |
2f714160 | 1449 | goto try_again; |
fcc00621 NH |
1450 | } |
1451 | ret = -EIO; | |
0ed950d1 | 1452 | goto out; |
2f714160 | 1453 | } |
0ed950d1 NH |
1454 | } |
1455 | ||
bf6445bc | 1456 | if (PageHuge(p) || HWPoisonHandlable(p, flags)) { |
0ed950d1 | 1457 | ret = 1; |
2f714160 | 1458 | } else { |
0ed950d1 NH |
1459 | /* |
1460 | * A page we cannot handle. Check whether we can turn | |
1461 | * it into something we can handle. | |
1462 | */ | |
1463 | if (pass++ < 3) { | |
2f714160 | 1464 | put_page(p); |
d0505e9f | 1465 | shake_page(p); |
0ed950d1 NH |
1466 | count_increased = false; |
1467 | goto try_again; | |
2f714160 | 1468 | } |
0ed950d1 NH |
1469 | put_page(p); |
1470 | ret = -EIO; | |
17e395b6 | 1471 | } |
0ed950d1 | 1472 | out: |
941ca063 | 1473 | if (ret == -EIO) |
96f96763 | 1474 | pr_err("%#lx: unhandlable page.\n", page_to_pfn(p)); |
941ca063 | 1475 | |
17e395b6 OS |
1476 | return ret; |
1477 | } | |
1478 | ||
bf181c58 NH |
1479 | static int __get_unpoison_page(struct page *page) |
1480 | { | |
04bac040 | 1481 | struct folio *folio = page_folio(page); |
bf181c58 NH |
1482 | int ret = 0; |
1483 | bool hugetlb = false; | |
1484 | ||
04bac040 | 1485 | ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, true); |
d31155b8 ML |
1486 | if (hugetlb) { |
1487 | /* Make sure hugetlb demotion did not happen from under us. */ | |
1488 | if (folio == page_folio(page)) | |
1489 | return ret; | |
1490 | if (ret > 0) | |
1491 | folio_put(folio); | |
1492 | } | |
bf181c58 NH |
1493 | |
1494 | /* | |
1495 | * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison, | |
1496 | * but also isolated from buddy freelist, so need to identify the | |
1497 | * state and have to cancel both operations to unpoison. | |
1498 | */ | |
1499 | if (PageHWPoisonTakenOff(page)) | |
1500 | return -EHWPOISON; | |
1501 | ||
1502 | return get_page_unless_zero(page) ? 1 : 0; | |
1503 | } | |
1504 | ||
0ed950d1 NH |
1505 | /** |
1506 | * get_hwpoison_page() - Get refcount for memory error handling | |
1507 | * @p: Raw error page (hit by memory error) | |
1508 | * @flags: Flags controlling behavior of error handling | |
1509 | * | |
1510 | * get_hwpoison_page() takes a page refcount of an error page to handle memory | |
1511 | * error on it, after checking that the error page is in a well-defined state | |
0b8f0d87 | 1512 | * (defined as a page-type we can successfully handle the memory error on it, |
0ed950d1 NH |
1513 | * such as LRU page and hugetlb page). |
1514 | * | |
1515 | * Memory error handling could be triggered at any time on any type of page, | |
1516 | * so it's prone to race with typical memory management lifecycle (like | |
1517 | * allocation and free). So to avoid such races, get_hwpoison_page() takes | |
1518 | * extra care for the error page's state (as done in __get_hwpoison_page()), | |
1519 | * and has some retry logic in get_any_page(). | |
1520 | * | |
bf181c58 NH |
1521 | * When called from unpoison_memory(), the caller should already ensure that |
1522 | * the given page has PG_hwpoison. So it's never reused for other page | |
1523 | * allocations, and __get_unpoison_page() never races with them. | |
1524 | * | |
0ed950d1 NH |
1525 | * Return: 0 on failure, |
1526 | * 1 on success for in-use pages in a well-defined state, | |
1527 | * -EIO for pages on which we can not handle memory errors, | |
1528 | * -EBUSY when get_hwpoison_page() has raced with page lifecycle | |
bf181c58 NH |
1529 | * operations like allocation and free, |
1530 | * -EHWPOISON when the page is hwpoisoned and taken off from buddy. | |
0ed950d1 NH |
1531 | */ |
1532 | static int get_hwpoison_page(struct page *p, unsigned long flags) | |
2f714160 OS |
1533 | { |
1534 | int ret; | |
1535 | ||
1536 | zone_pcp_disable(page_zone(p)); | |
bf181c58 NH |
1537 | if (flags & MF_UNPOISON) |
1538 | ret = __get_unpoison_page(p); | |
1539 | else | |
1540 | ret = get_any_page(p, flags); | |
2f714160 OS |
1541 | zone_pcp_enable(page_zone(p)); |
1542 | ||
1543 | return ret; | |
1544 | } | |
1545 | ||
6a46079c AK |
1546 | /* |
1547 | * Do all that is necessary to remove user space mappings. Unmap | |
1548 | * the pages and send SIGBUS to the processes if the data was dirty. | |
1549 | */ | |
666e5a40 | 1550 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
ed8c2f49 | 1551 | int flags, struct page *hpage) |
6a46079c | 1552 | { |
869f7ee6 | 1553 | struct folio *folio = page_folio(hpage); |
6da6b1d4 | 1554 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC | TTU_HWPOISON; |
6a46079c AK |
1555 | struct address_space *mapping; |
1556 | LIST_HEAD(tokill); | |
1fb08ac6 | 1557 | bool unmap_success; |
0792a4a6 | 1558 | int forcekill; |
286c469a | 1559 | bool mlocked = PageMlocked(hpage); |
6a46079c | 1560 | |
93a9eb39 NH |
1561 | /* |
1562 | * Here we are interested only in user-mapped pages, so skip any | |
1563 | * other types of pages. | |
1564 | */ | |
7a8817f2 | 1565 | if (PageReserved(p) || PageSlab(p) || PageTable(p) || PageOffline(p)) |
666e5a40 | 1566 | return true; |
93a9eb39 | 1567 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 1568 | return true; |
6a46079c | 1569 | |
6a46079c AK |
1570 | /* |
1571 | * This check implies we don't kill processes if their pages | |
1572 | * are in the swap cache early. Those are always late kills. | |
1573 | */ | |
7af446a8 | 1574 | if (!page_mapped(hpage)) |
666e5a40 | 1575 | return true; |
1668bfd5 | 1576 | |
6a46079c | 1577 | if (PageSwapCache(p)) { |
96f96763 | 1578 | pr_err("%#lx: keeping poisoned page in swap cache\n", pfn); |
6da6b1d4 | 1579 | ttu &= ~TTU_HWPOISON; |
6a46079c AK |
1580 | } |
1581 | ||
1582 | /* | |
1583 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1584 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1585 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1586 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1587 | */ |
7af446a8 | 1588 | mapping = page_mapping(hpage); |
6751ed65 | 1589 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
f56753ac | 1590 | mapping_can_writeback(mapping)) { |
7af446a8 NH |
1591 | if (page_mkclean(hpage)) { |
1592 | SetPageDirty(hpage); | |
6a46079c | 1593 | } else { |
6da6b1d4 | 1594 | ttu &= ~TTU_HWPOISON; |
96f96763 | 1595 | pr_info("%#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1596 | pfn); |
1597 | } | |
1598 | } | |
1599 | ||
1600 | /* | |
1601 | * First collect all the processes that have the page | |
1602 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1603 | * because ttu takes the rmap data structures down. | |
6a46079c | 1604 | */ |
0792a4a6 | 1605 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1606 | |
357670f7 ML |
1607 | if (PageHuge(hpage) && !PageAnon(hpage)) { |
1608 | /* | |
1609 | * For hugetlb pages in shared mappings, try_to_unmap | |
1610 | * could potentially call huge_pmd_unshare. Because of | |
1611 | * this, take semaphore in write mode here and set | |
1612 | * TTU_RMAP_LOCKED to indicate we have taken the lock | |
1613 | * at this higher level. | |
1614 | */ | |
1615 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1616 | if (mapping) { | |
9030fb0b | 1617 | try_to_unmap(folio, ttu|TTU_RMAP_LOCKED); |
357670f7 ML |
1618 | i_mmap_unlock_write(mapping); |
1619 | } else | |
96f96763 | 1620 | pr_info("%#lx: could not lock mapping for mapped huge page\n", pfn); |
c0d0381a | 1621 | } else { |
9030fb0b | 1622 | try_to_unmap(folio, ttu); |
c0d0381a | 1623 | } |
1fb08ac6 YS |
1624 | |
1625 | unmap_success = !page_mapped(hpage); | |
666e5a40 | 1626 | if (!unmap_success) |
96f96763 | 1627 | pr_err("%#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1628 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1629 | |
286c469a NH |
1630 | /* |
1631 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1632 | * shake_page() again to ensure that it's flushed. | |
1633 | */ | |
1634 | if (mlocked) | |
d0505e9f | 1635 | shake_page(hpage); |
286c469a | 1636 | |
6a46079c AK |
1637 | /* |
1638 | * Now that the dirty bit has been propagated to the | |
1639 | * struct page and all unmaps done we can decide if | |
1640 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1641 | * was dirty or the process is not restartable, |
1642 | * otherwise the tokill list is merely | |
6a46079c AK |
1643 | * freed. When there was a problem unmapping earlier |
1644 | * use a more force-full uncatchable kill to prevent | |
1645 | * any accesses to the poisoned memory. | |
1646 | */ | |
0792a4a6 ML |
1647 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL) || |
1648 | !unmap_success; | |
ae1139ec | 1649 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1650 | |
666e5a40 | 1651 | return unmap_success; |
6a46079c AK |
1652 | } |
1653 | ||
0348d2eb NH |
1654 | static int identify_page_state(unsigned long pfn, struct page *p, |
1655 | unsigned long page_flags) | |
761ad8d7 NH |
1656 | { |
1657 | struct page_state *ps; | |
0348d2eb NH |
1658 | |
1659 | /* | |
1660 | * The first check uses the current page flags which may not have any | |
1661 | * relevant information. The second check with the saved page flags is | |
1662 | * carried out only if the first check can't determine the page status. | |
1663 | */ | |
1664 | for (ps = error_states;; ps++) | |
1665 | if ((p->flags & ps->mask) == ps->res) | |
1666 | break; | |
1667 | ||
1668 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1669 | ||
1670 | if (!ps->mask) | |
1671 | for (ps = error_states;; ps++) | |
1672 | if ((page_flags & ps->mask) == ps->res) | |
1673 | break; | |
1674 | return page_action(ps, p, pfn); | |
1675 | } | |
1676 | ||
2ace36f0 | 1677 | static int try_to_split_thp_page(struct page *page) |
694bf0b0 | 1678 | { |
2ace36f0 KW |
1679 | int ret; |
1680 | ||
694bf0b0 | 1681 | lock_page(page); |
2ace36f0 KW |
1682 | ret = split_huge_page(page); |
1683 | unlock_page(page); | |
694bf0b0 | 1684 | |
2ace36f0 | 1685 | if (unlikely(ret)) |
694bf0b0 | 1686 | put_page(page); |
694bf0b0 | 1687 | |
2ace36f0 | 1688 | return ret; |
694bf0b0 OS |
1689 | } |
1690 | ||
00cc790e SR |
1691 | static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn, |
1692 | struct address_space *mapping, pgoff_t index, int flags) | |
1693 | { | |
1694 | struct to_kill *tk; | |
1695 | unsigned long size = 0; | |
1696 | ||
1697 | list_for_each_entry(tk, to_kill, nd) | |
1698 | if (tk->size_shift) | |
1699 | size = max(size, 1UL << tk->size_shift); | |
1700 | ||
1701 | if (size) { | |
1702 | /* | |
1703 | * Unmap the largest mapping to avoid breaking up device-dax | |
1704 | * mappings which are constant size. The actual size of the | |
1705 | * mapping being torn down is communicated in siginfo, see | |
1706 | * kill_proc() | |
1707 | */ | |
1708 | loff_t start = (index << PAGE_SHIFT) & ~(size - 1); | |
1709 | ||
1710 | unmap_mapping_range(mapping, start, size, 0); | |
1711 | } | |
1712 | ||
1713 | kill_procs(to_kill, flags & MF_MUST_KILL, false, pfn, flags); | |
1714 | } | |
1715 | ||
1716 | static int mf_generic_kill_procs(unsigned long long pfn, int flags, | |
1717 | struct dev_pagemap *pgmap) | |
1718 | { | |
1719 | struct page *page = pfn_to_page(pfn); | |
1720 | LIST_HEAD(to_kill); | |
1721 | dax_entry_t cookie; | |
1722 | int rc = 0; | |
1723 | ||
1724 | /* | |
1725 | * Pages instantiated by device-dax (not filesystem-dax) | |
1726 | * may be compound pages. | |
1727 | */ | |
1728 | page = compound_head(page); | |
1729 | ||
1730 | /* | |
1731 | * Prevent the inode from being freed while we are interrogating | |
1732 | * the address_space, typically this would be handled by | |
1733 | * lock_page(), but dax pages do not use the page lock. This | |
1734 | * also prevents changes to the mapping of this pfn until | |
1735 | * poison signaling is complete. | |
1736 | */ | |
1737 | cookie = dax_lock_page(page); | |
1738 | if (!cookie) | |
1739 | return -EBUSY; | |
1740 | ||
1741 | if (hwpoison_filter(page)) { | |
1742 | rc = -EOPNOTSUPP; | |
1743 | goto unlock; | |
1744 | } | |
1745 | ||
1746 | switch (pgmap->type) { | |
1747 | case MEMORY_DEVICE_PRIVATE: | |
1748 | case MEMORY_DEVICE_COHERENT: | |
1749 | /* | |
1750 | * TODO: Handle device pages which may need coordination | |
1751 | * with device-side memory. | |
1752 | */ | |
1753 | rc = -ENXIO; | |
1754 | goto unlock; | |
1755 | default: | |
1756 | break; | |
1757 | } | |
1758 | ||
1759 | /* | |
1760 | * Use this flag as an indication that the dax page has been | |
1761 | * remapped UC to prevent speculative consumption of poison. | |
1762 | */ | |
1763 | SetPageHWPoison(page); | |
1764 | ||
1765 | /* | |
1766 | * Unlike System-RAM there is no possibility to swap in a | |
1767 | * different physical page at a given virtual address, so all | |
1768 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1769 | * SIGBUS (i.e. MF_MUST_KILL) | |
1770 | */ | |
1771 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1772 | collect_procs(page, &to_kill, true); | |
1773 | ||
1774 | unmap_and_kill(&to_kill, pfn, page->mapping, page->index, flags); | |
1775 | unlock: | |
1776 | dax_unlock_page(page, cookie); | |
1777 | return rc; | |
1778 | } | |
1779 | ||
c36e2024 SR |
1780 | #ifdef CONFIG_FS_DAX |
1781 | /** | |
1782 | * mf_dax_kill_procs - Collect and kill processes who are using this file range | |
1783 | * @mapping: address_space of the file in use | |
1784 | * @index: start pgoff of the range within the file | |
1785 | * @count: length of the range, in unit of PAGE_SIZE | |
1786 | * @mf_flags: memory failure flags | |
1787 | */ | |
1788 | int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index, | |
1789 | unsigned long count, int mf_flags) | |
1790 | { | |
1791 | LIST_HEAD(to_kill); | |
1792 | dax_entry_t cookie; | |
1793 | struct page *page; | |
1794 | size_t end = index + count; | |
1795 | ||
1796 | mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1797 | ||
1798 | for (; index < end; index++) { | |
1799 | page = NULL; | |
1800 | cookie = dax_lock_mapping_entry(mapping, index, &page); | |
1801 | if (!cookie) | |
1802 | return -EBUSY; | |
1803 | if (!page) | |
1804 | goto unlock; | |
1805 | ||
1806 | SetPageHWPoison(page); | |
1807 | ||
1808 | collect_procs_fsdax(page, mapping, index, &to_kill); | |
1809 | unmap_and_kill(&to_kill, page_to_pfn(page), mapping, | |
1810 | index, mf_flags); | |
1811 | unlock: | |
1812 | dax_unlock_mapping_entry(mapping, index, cookie); | |
1813 | } | |
1814 | return 0; | |
1815 | } | |
1816 | EXPORT_SYMBOL_GPL(mf_dax_kill_procs); | |
1817 | #endif /* CONFIG_FS_DAX */ | |
1818 | ||
161df60e | 1819 | #ifdef CONFIG_HUGETLB_PAGE |
b79f8eb4 | 1820 | |
161df60e NH |
1821 | /* |
1822 | * Struct raw_hwp_page represents information about "raw error page", | |
dad6a5eb | 1823 | * constructing singly linked list from ->_hugetlb_hwpoison field of folio. |
161df60e NH |
1824 | */ |
1825 | struct raw_hwp_page { | |
1826 | struct llist_node node; | |
1827 | struct page *page; | |
1828 | }; | |
1829 | ||
b02e7582 | 1830 | static inline struct llist_head *raw_hwp_list_head(struct folio *folio) |
161df60e | 1831 | { |
b02e7582 | 1832 | return (struct llist_head *)&folio->_hugetlb_hwpoison; |
161df60e NH |
1833 | } |
1834 | ||
b79f8eb4 JY |
1835 | bool is_raw_hwpoison_page_in_hugepage(struct page *page) |
1836 | { | |
1837 | struct llist_head *raw_hwp_head; | |
1838 | struct raw_hwp_page *p; | |
1839 | struct folio *folio = page_folio(page); | |
1840 | bool ret = false; | |
1841 | ||
1842 | if (!folio_test_hwpoison(folio)) | |
1843 | return false; | |
1844 | ||
1845 | if (!folio_test_hugetlb(folio)) | |
1846 | return PageHWPoison(page); | |
1847 | ||
1848 | /* | |
1849 | * When RawHwpUnreliable is set, kernel lost track of which subpages | |
1850 | * are HWPOISON. So return as if ALL subpages are HWPOISONed. | |
1851 | */ | |
1852 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) | |
1853 | return true; | |
1854 | ||
1855 | mutex_lock(&mf_mutex); | |
1856 | ||
1857 | raw_hwp_head = raw_hwp_list_head(folio); | |
1858 | llist_for_each_entry(p, raw_hwp_head->first, node) { | |
1859 | if (page == p->page) { | |
1860 | ret = true; | |
1861 | break; | |
1862 | } | |
1863 | } | |
1864 | ||
1865 | mutex_unlock(&mf_mutex); | |
1866 | ||
1867 | return ret; | |
1868 | } | |
1869 | ||
0858b5eb | 1870 | static unsigned long __folio_free_raw_hwp(struct folio *folio, bool move_flag) |
161df60e | 1871 | { |
6379693e ML |
1872 | struct llist_node *head; |
1873 | struct raw_hwp_page *p, *next; | |
ac5fcde0 | 1874 | unsigned long count = 0; |
161df60e | 1875 | |
9e130c4b | 1876 | head = llist_del_all(raw_hwp_list_head(folio)); |
6379693e | 1877 | llist_for_each_entry_safe(p, next, head, node) { |
ac5fcde0 NH |
1878 | if (move_flag) |
1879 | SetPageHWPoison(p->page); | |
5033091d NH |
1880 | else |
1881 | num_poisoned_pages_sub(page_to_pfn(p->page), 1); | |
161df60e | 1882 | kfree(p); |
ac5fcde0 | 1883 | count++; |
161df60e | 1884 | } |
ac5fcde0 | 1885 | return count; |
161df60e NH |
1886 | } |
1887 | ||
595dd818 | 1888 | static int folio_set_hugetlb_hwpoison(struct folio *folio, struct page *page) |
161df60e NH |
1889 | { |
1890 | struct llist_head *head; | |
1891 | struct raw_hwp_page *raw_hwp; | |
6379693e | 1892 | struct raw_hwp_page *p, *next; |
595dd818 | 1893 | int ret = folio_test_set_hwpoison(folio) ? -EHWPOISON : 0; |
161df60e NH |
1894 | |
1895 | /* | |
1896 | * Once the hwpoison hugepage has lost reliable raw error info, | |
1897 | * there is little meaning to keep additional error info precisely, | |
1898 | * so skip to add additional raw error info. | |
1899 | */ | |
b02e7582 | 1900 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) |
161df60e | 1901 | return -EHWPOISON; |
b02e7582 | 1902 | head = raw_hwp_list_head(folio); |
6379693e | 1903 | llist_for_each_entry_safe(p, next, head->first, node) { |
161df60e NH |
1904 | if (p->page == page) |
1905 | return -EHWPOISON; | |
1906 | } | |
1907 | ||
1908 | raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC); | |
1909 | if (raw_hwp) { | |
1910 | raw_hwp->page = page; | |
1911 | llist_add(&raw_hwp->node, head); | |
1912 | /* the first error event will be counted in action_result(). */ | |
1913 | if (ret) | |
a46c9304 | 1914 | num_poisoned_pages_inc(page_to_pfn(page)); |
161df60e NH |
1915 | } else { |
1916 | /* | |
1917 | * Failed to save raw error info. We no longer trace all | |
1918 | * hwpoisoned subpages, and we need refuse to free/dissolve | |
1919 | * this hwpoisoned hugepage. | |
1920 | */ | |
b02e7582 | 1921 | folio_set_hugetlb_raw_hwp_unreliable(folio); |
161df60e | 1922 | /* |
b02e7582 | 1923 | * Once hugetlb_raw_hwp_unreliable is set, raw_hwp_page is not |
161df60e NH |
1924 | * used any more, so free it. |
1925 | */ | |
0858b5eb | 1926 | __folio_free_raw_hwp(folio, false); |
161df60e NH |
1927 | } |
1928 | return ret; | |
1929 | } | |
1930 | ||
9637d7df | 1931 | static unsigned long folio_free_raw_hwp(struct folio *folio, bool move_flag) |
ac5fcde0 NH |
1932 | { |
1933 | /* | |
9637d7df | 1934 | * hugetlb_vmemmap_optimized hugepages can't be freed because struct |
ac5fcde0 NH |
1935 | * pages for tail pages are required but they don't exist. |
1936 | */ | |
9637d7df | 1937 | if (move_flag && folio_test_hugetlb_vmemmap_optimized(folio)) |
ac5fcde0 NH |
1938 | return 0; |
1939 | ||
1940 | /* | |
9637d7df | 1941 | * hugetlb_raw_hwp_unreliable hugepages shouldn't be unpoisoned by |
ac5fcde0 NH |
1942 | * definition. |
1943 | */ | |
9637d7df | 1944 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) |
ac5fcde0 NH |
1945 | return 0; |
1946 | ||
0858b5eb | 1947 | return __folio_free_raw_hwp(folio, move_flag); |
ac5fcde0 NH |
1948 | } |
1949 | ||
2ff6cece | 1950 | void folio_clear_hugetlb_hwpoison(struct folio *folio) |
161df60e | 1951 | { |
2ff6cece | 1952 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) |
161df60e | 1953 | return; |
92a025a7 ML |
1954 | if (folio_test_hugetlb_vmemmap_optimized(folio)) |
1955 | return; | |
2ff6cece | 1956 | folio_clear_hwpoison(folio); |
9637d7df | 1957 | folio_free_raw_hwp(folio, true); |
161df60e NH |
1958 | } |
1959 | ||
405ce051 NH |
1960 | /* |
1961 | * Called from hugetlb code with hugetlb_lock held. | |
1962 | * | |
1963 | * Return values: | |
1964 | * 0 - free hugepage | |
1965 | * 1 - in-use hugepage | |
1966 | * 2 - not a hugepage | |
1967 | * -EBUSY - the hugepage is busy (try to retry) | |
1968 | * -EHWPOISON - the hugepage is already hwpoisoned | |
1969 | */ | |
e591ef7d NH |
1970 | int __get_huge_page_for_hwpoison(unsigned long pfn, int flags, |
1971 | bool *migratable_cleared) | |
405ce051 NH |
1972 | { |
1973 | struct page *page = pfn_to_page(pfn); | |
4c110ec9 | 1974 | struct folio *folio = page_folio(page); |
405ce051 NH |
1975 | int ret = 2; /* fallback to normal page handling */ |
1976 | bool count_increased = false; | |
1977 | ||
4c110ec9 | 1978 | if (!folio_test_hugetlb(folio)) |
405ce051 NH |
1979 | goto out; |
1980 | ||
1981 | if (flags & MF_COUNT_INCREASED) { | |
1982 | ret = 1; | |
1983 | count_increased = true; | |
4c110ec9 | 1984 | } else if (folio_test_hugetlb_freed(folio)) { |
b283d983 | 1985 | ret = 0; |
4c110ec9 SK |
1986 | } else if (folio_test_hugetlb_migratable(folio)) { |
1987 | ret = folio_try_get(folio); | |
405ce051 NH |
1988 | if (ret) |
1989 | count_increased = true; | |
1990 | } else { | |
1991 | ret = -EBUSY; | |
38f6d293 NH |
1992 | if (!(flags & MF_NO_RETRY)) |
1993 | goto out; | |
405ce051 NH |
1994 | } |
1995 | ||
595dd818 | 1996 | if (folio_set_hugetlb_hwpoison(folio, page)) { |
405ce051 NH |
1997 | ret = -EHWPOISON; |
1998 | goto out; | |
1999 | } | |
2000 | ||
e591ef7d | 2001 | /* |
4c110ec9 | 2002 | * Clearing hugetlb_migratable for hwpoisoned hugepages to prevent them |
e591ef7d NH |
2003 | * from being migrated by memory hotremove. |
2004 | */ | |
4c110ec9 SK |
2005 | if (count_increased && folio_test_hugetlb_migratable(folio)) { |
2006 | folio_clear_hugetlb_migratable(folio); | |
e591ef7d NH |
2007 | *migratable_cleared = true; |
2008 | } | |
2009 | ||
405ce051 NH |
2010 | return ret; |
2011 | out: | |
2012 | if (count_increased) | |
4c110ec9 | 2013 | folio_put(folio); |
405ce051 NH |
2014 | return ret; |
2015 | } | |
2016 | ||
405ce051 NH |
2017 | /* |
2018 | * Taking refcount of hugetlb pages needs extra care about race conditions | |
2019 | * with basic operations like hugepage allocation/free/demotion. | |
2020 | * So some of prechecks for hwpoison (pinning, and testing/setting | |
2021 | * PageHWPoison) should be done in single hugetlb_lock range. | |
2022 | */ | |
2023 | static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) | |
0348d2eb | 2024 | { |
761ad8d7 | 2025 | int res; |
405ce051 | 2026 | struct page *p = pfn_to_page(pfn); |
bc1cfde1 | 2027 | struct folio *folio; |
761ad8d7 | 2028 | unsigned long page_flags; |
e591ef7d | 2029 | bool migratable_cleared = false; |
761ad8d7 | 2030 | |
405ce051 NH |
2031 | *hugetlb = 1; |
2032 | retry: | |
e591ef7d | 2033 | res = get_huge_page_for_hwpoison(pfn, flags, &migratable_cleared); |
405ce051 NH |
2034 | if (res == 2) { /* fallback to normal page handling */ |
2035 | *hugetlb = 0; | |
2036 | return 0; | |
2037 | } else if (res == -EHWPOISON) { | |
96f96763 | 2038 | pr_err("%#lx: already hardware poisoned\n", pfn); |
405ce051 | 2039 | if (flags & MF_ACTION_REQUIRED) { |
bc1cfde1 SK |
2040 | folio = page_folio(p); |
2041 | res = kill_accessing_process(current, folio_pfn(folio), flags); | |
405ce051 NH |
2042 | } |
2043 | return res; | |
2044 | } else if (res == -EBUSY) { | |
38f6d293 NH |
2045 | if (!(flags & MF_NO_RETRY)) { |
2046 | flags |= MF_NO_RETRY; | |
405ce051 NH |
2047 | goto retry; |
2048 | } | |
b66d00df | 2049 | return action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); |
761ad8d7 NH |
2050 | } |
2051 | ||
bc1cfde1 SK |
2052 | folio = page_folio(p); |
2053 | folio_lock(folio); | |
405ce051 NH |
2054 | |
2055 | if (hwpoison_filter(p)) { | |
2ff6cece | 2056 | folio_clear_hugetlb_hwpoison(folio); |
e591ef7d | 2057 | if (migratable_cleared) |
bc1cfde1 SK |
2058 | folio_set_hugetlb_migratable(folio); |
2059 | folio_unlock(folio); | |
f36a5543 | 2060 | if (res == 1) |
bc1cfde1 | 2061 | folio_put(folio); |
f36a5543 | 2062 | return -EOPNOTSUPP; |
405ce051 NH |
2063 | } |
2064 | ||
405ce051 NH |
2065 | /* |
2066 | * Handling free hugepage. The possible race with hugepage allocation | |
2067 | * or demotion can be prevented by PageHWPoison flag. | |
2068 | */ | |
2069 | if (res == 0) { | |
bc1cfde1 | 2070 | folio_unlock(folio); |
ceaf8fbe | 2071 | if (__page_handle_poison(p) >= 0) { |
405ce051 NH |
2072 | page_ref_inc(p); |
2073 | res = MF_RECOVERED; | |
ceaf8fbe NH |
2074 | } else { |
2075 | res = MF_FAILED; | |
761ad8d7 | 2076 | } |
b66d00df | 2077 | return action_result(pfn, MF_MSG_FREE_HUGE, res); |
761ad8d7 NH |
2078 | } |
2079 | ||
bc1cfde1 | 2080 | page_flags = folio->flags; |
761ad8d7 | 2081 | |
bc1cfde1 SK |
2082 | if (!hwpoison_user_mappings(p, pfn, flags, &folio->page)) { |
2083 | folio_unlock(folio); | |
b66d00df | 2084 | return action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
761ad8d7 NH |
2085 | } |
2086 | ||
ea6d0630 | 2087 | return identify_page_state(pfn, p, page_flags); |
761ad8d7 | 2088 | } |
00cc790e | 2089 | |
405ce051 NH |
2090 | #else |
2091 | static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) | |
2092 | { | |
2093 | return 0; | |
2094 | } | |
00cc790e | 2095 | |
9637d7df | 2096 | static inline unsigned long folio_free_raw_hwp(struct folio *folio, bool flag) |
ac5fcde0 NH |
2097 | { |
2098 | return 0; | |
2099 | } | |
00cc790e | 2100 | #endif /* CONFIG_HUGETLB_PAGE */ |
761ad8d7 | 2101 | |
b5f1fc98 KW |
2102 | /* Drop the extra refcount in case we come from madvise() */ |
2103 | static void put_ref_page(unsigned long pfn, int flags) | |
2104 | { | |
2105 | struct page *page; | |
2106 | ||
2107 | if (!(flags & MF_COUNT_INCREASED)) | |
2108 | return; | |
2109 | ||
2110 | page = pfn_to_page(pfn); | |
2111 | if (page) | |
2112 | put_page(page); | |
2113 | } | |
2114 | ||
6100e34b DW |
2115 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
2116 | struct dev_pagemap *pgmap) | |
2117 | { | |
00cc790e | 2118 | int rc = -ENXIO; |
6100e34b | 2119 | |
34dc45be | 2120 | /* device metadata space is not recoverable */ |
00cc790e | 2121 | if (!pgmap_pfn_valid(pgmap, pfn)) |
34dc45be | 2122 | goto out; |
61e28cf0 | 2123 | |
6100e34b | 2124 | /* |
33a8f7f2 SR |
2125 | * Call driver's implementation to handle the memory failure, otherwise |
2126 | * fall back to generic handler. | |
6100e34b | 2127 | */ |
65d3440e | 2128 | if (pgmap_has_memory_failure(pgmap)) { |
33a8f7f2 | 2129 | rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags); |
6100e34b | 2130 | /* |
33a8f7f2 SR |
2131 | * Fall back to generic handler too if operation is not |
2132 | * supported inside the driver/device/filesystem. | |
6100e34b | 2133 | */ |
33a8f7f2 SR |
2134 | if (rc != -EOPNOTSUPP) |
2135 | goto out; | |
6100e34b DW |
2136 | } |
2137 | ||
00cc790e | 2138 | rc = mf_generic_kill_procs(pfn, flags, pgmap); |
6100e34b DW |
2139 | out: |
2140 | /* drop pgmap ref acquired in caller */ | |
2141 | put_dev_pagemap(pgmap); | |
80ee7cb2 ML |
2142 | if (rc != -EOPNOTSUPP) |
2143 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
6100e34b DW |
2144 | return rc; |
2145 | } | |
2146 | ||
cd42f4a3 TL |
2147 | /** |
2148 | * memory_failure - Handle memory failure of a page. | |
2149 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
2150 | * @flags: fine tune action taken |
2151 | * | |
2152 | * This function is called by the low level machine check code | |
2153 | * of an architecture when it detects hardware memory corruption | |
2154 | * of a page. It tries its best to recover, which includes | |
2155 | * dropping pages, killing processes etc. | |
2156 | * | |
2157 | * The function is primarily of use for corruptions that | |
2158 | * happen outside the current execution context (e.g. when | |
2159 | * detected by a background scrubber) | |
2160 | * | |
2161 | * Must run in process context (e.g. a work queue) with interrupts | |
5885c6a6 | 2162 | * enabled and no spinlocks held. |
d1fe111f | 2163 | * |
2164 | * Return: 0 for successfully handled the memory error, | |
9113eaf3 | 2165 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event, |
d1fe111f | 2166 | * < 0(except -EOPNOTSUPP) on failure. |
cd42f4a3 | 2167 | */ |
83b57531 | 2168 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 2169 | { |
6a46079c | 2170 | struct page *p; |
7af446a8 | 2171 | struct page *hpage; |
6100e34b | 2172 | struct dev_pagemap *pgmap; |
171936dd | 2173 | int res = 0; |
524fca1e | 2174 | unsigned long page_flags; |
a8b2c2ce | 2175 | bool retry = true; |
405ce051 | 2176 | int hugetlb = 0; |
6a46079c AK |
2177 | |
2178 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 2179 | panic("Memory failure on page %lx", pfn); |
6a46079c | 2180 | |
03b122da TL |
2181 | mutex_lock(&mf_mutex); |
2182 | ||
67f22ba7 | 2183 | if (!(flags & MF_SW_SIMULATED)) |
2184 | hw_memory_failure = true; | |
2185 | ||
96c804a6 DH |
2186 | p = pfn_to_online_page(pfn); |
2187 | if (!p) { | |
03b122da TL |
2188 | res = arch_memory_failure(pfn, flags); |
2189 | if (res == 0) | |
2190 | goto unlock_mutex; | |
2191 | ||
96c804a6 DH |
2192 | if (pfn_valid(pfn)) { |
2193 | pgmap = get_dev_pagemap(pfn, NULL); | |
d51b6846 | 2194 | put_ref_page(pfn, flags); |
03b122da TL |
2195 | if (pgmap) { |
2196 | res = memory_failure_dev_pagemap(pfn, flags, | |
2197 | pgmap); | |
2198 | goto unlock_mutex; | |
2199 | } | |
96c804a6 | 2200 | } |
96f96763 | 2201 | pr_err("%#lx: memory outside kernel control\n", pfn); |
03b122da TL |
2202 | res = -ENXIO; |
2203 | goto unlock_mutex; | |
6a46079c AK |
2204 | } |
2205 | ||
a8b2c2ce | 2206 | try_again: |
405ce051 NH |
2207 | res = try_memory_failure_hugetlb(pfn, flags, &hugetlb); |
2208 | if (hugetlb) | |
171936dd | 2209 | goto unlock_mutex; |
171936dd | 2210 | |
6a46079c | 2211 | if (TestSetPageHWPoison(p)) { |
96f96763 | 2212 | pr_err("%#lx: already hardware poisoned\n", pfn); |
47af12ba | 2213 | res = -EHWPOISON; |
a3f5d80e NH |
2214 | if (flags & MF_ACTION_REQUIRED) |
2215 | res = kill_accessing_process(current, pfn, flags); | |
f361e246 NH |
2216 | if (flags & MF_COUNT_INCREASED) |
2217 | put_page(p); | |
171936dd | 2218 | goto unlock_mutex; |
6a46079c AK |
2219 | } |
2220 | ||
6a46079c AK |
2221 | /* |
2222 | * We need/can do nothing about count=0 pages. | |
2223 | * 1) it's a free page, and therefore in safe hand: | |
9cf28191 | 2224 | * check_new_page() will be the gate keeper. |
761ad8d7 | 2225 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
2226 | * Implies some kernel user: cannot stop them from |
2227 | * R/W the page; let's pray that the page has been | |
2228 | * used and will be freed some time later. | |
2229 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 2230 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 2231 | */ |
0ed950d1 NH |
2232 | if (!(flags & MF_COUNT_INCREASED)) { |
2233 | res = get_hwpoison_page(p, flags); | |
2234 | if (!res) { | |
2235 | if (is_free_buddy_page(p)) { | |
2236 | if (take_page_off_buddy(p)) { | |
2237 | page_ref_inc(p); | |
2238 | res = MF_RECOVERED; | |
2239 | } else { | |
2240 | /* We lost the race, try again */ | |
2241 | if (retry) { | |
2242 | ClearPageHWPoison(p); | |
0ed950d1 NH |
2243 | retry = false; |
2244 | goto try_again; | |
2245 | } | |
2246 | res = MF_FAILED; | |
a8b2c2ce | 2247 | } |
b66d00df | 2248 | res = action_result(pfn, MF_MSG_BUDDY, res); |
0ed950d1 | 2249 | } else { |
b66d00df | 2250 | res = action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); |
a8b2c2ce | 2251 | } |
0ed950d1 NH |
2252 | goto unlock_mutex; |
2253 | } else if (res < 0) { | |
b66d00df | 2254 | res = action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); |
0ed950d1 | 2255 | goto unlock_mutex; |
8d22ba1b | 2256 | } |
6a46079c AK |
2257 | } |
2258 | ||
a363d122 | 2259 | hpage = compound_head(p); |
761ad8d7 | 2260 | if (PageTransHuge(hpage)) { |
eac96c3e YS |
2261 | /* |
2262 | * The flag must be set after the refcount is bumped | |
2263 | * otherwise it may race with THP split. | |
2264 | * And the flag can't be set in get_hwpoison_page() since | |
2265 | * it is called by soft offline too and it is just called | |
5885c6a6 | 2266 | * for !MF_COUNT_INCREASED. So here seems to be the best |
eac96c3e YS |
2267 | * place. |
2268 | * | |
2269 | * Don't need care about the above error handling paths for | |
2270 | * get_hwpoison_page() since they handle either free page | |
2271 | * or unhandlable page. The refcount is bumped iff the | |
2272 | * page is a valid handlable page. | |
2273 | */ | |
2274 | SetPageHasHWPoisoned(hpage); | |
2ace36f0 | 2275 | if (try_to_split_thp_page(p) < 0) { |
b66d00df | 2276 | res = action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED); |
171936dd | 2277 | goto unlock_mutex; |
5d1fd5dc | 2278 | } |
415c64c1 | 2279 | VM_BUG_ON_PAGE(!page_count(p), p); |
415c64c1 NH |
2280 | } |
2281 | ||
e43c3afb WF |
2282 | /* |
2283 | * We ignore non-LRU pages for good reasons. | |
2284 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 2285 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
2286 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
2287 | * The check (unnecessarily) ignores LRU pages being isolated and | |
2288 | * walked by the page reclaim code, however that's not a big loss. | |
2289 | */ | |
d0505e9f | 2290 | shake_page(p); |
e43c3afb | 2291 | |
761ad8d7 | 2292 | lock_page(p); |
847ce401 | 2293 | |
f37d4298 | 2294 | /* |
75ee64b3 ML |
2295 | * We're only intended to deal with the non-Compound page here. |
2296 | * However, the page could have changed compound pages due to | |
2297 | * race window. If this happens, we could try again to hopefully | |
2298 | * handle the page next round. | |
f37d4298 | 2299 | */ |
75ee64b3 ML |
2300 | if (PageCompound(p)) { |
2301 | if (retry) { | |
e240ac52 | 2302 | ClearPageHWPoison(p); |
75ee64b3 ML |
2303 | unlock_page(p); |
2304 | put_page(p); | |
2305 | flags &= ~MF_COUNT_INCREASED; | |
2306 | retry = false; | |
2307 | goto try_again; | |
2308 | } | |
b66d00df | 2309 | res = action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
171936dd | 2310 | goto unlock_page; |
f37d4298 AK |
2311 | } |
2312 | ||
524fca1e NH |
2313 | /* |
2314 | * We use page flags to determine what action should be taken, but | |
2315 | * the flags can be modified by the error containment action. One | |
2316 | * example is an mlocked page, where PG_mlocked is cleared by | |
2317 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
2318 | * correctly, we save a copy of the page flags at this time. | |
2319 | */ | |
7d9d46ac | 2320 | page_flags = p->flags; |
524fca1e | 2321 | |
7c116f2b | 2322 | if (hwpoison_filter(p)) { |
2fe62e22 | 2323 | ClearPageHWPoison(p); |
761ad8d7 | 2324 | unlock_page(p); |
dd6e2402 | 2325 | put_page(p); |
d1fe111f | 2326 | res = -EOPNOTSUPP; |
171936dd | 2327 | goto unlock_mutex; |
7c116f2b | 2328 | } |
847ce401 | 2329 | |
e8675d29 | 2330 | /* |
e0650a41 | 2331 | * __munlock_folio() may clear a writeback page's LRU flag without |
e8675d29 | 2332 | * page_lock. We need wait writeback completion for this page or it |
2333 | * may trigger vfs BUG while evict inode. | |
2334 | */ | |
b04d3eeb | 2335 | if (!PageLRU(p) && !PageWriteback(p)) |
0bc1f8b0 CY |
2336 | goto identify_page_state; |
2337 | ||
6edd6cc6 NH |
2338 | /* |
2339 | * It's very difficult to mess with pages currently under IO | |
2340 | * and in many cases impossible, so we just avoid it here. | |
2341 | */ | |
6a46079c AK |
2342 | wait_on_page_writeback(p); |
2343 | ||
2344 | /* | |
2345 | * Now take care of user space mappings. | |
6ffcd825 | 2346 | * Abort on fail: __filemap_remove_folio() assumes unmapped page. |
6a46079c | 2347 | */ |
ed8c2f49 | 2348 | if (!hwpoison_user_mappings(p, pfn, flags, p)) { |
b66d00df | 2349 | res = action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
171936dd | 2350 | goto unlock_page; |
1668bfd5 | 2351 | } |
6a46079c AK |
2352 | |
2353 | /* | |
2354 | * Torn down by someone else? | |
2355 | */ | |
dc2a1cbf | 2356 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
b66d00df | 2357 | res = action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
171936dd | 2358 | goto unlock_page; |
6a46079c AK |
2359 | } |
2360 | ||
0bc1f8b0 | 2361 | identify_page_state: |
0348d2eb | 2362 | res = identify_page_state(pfn, p, page_flags); |
ea6d0630 NH |
2363 | mutex_unlock(&mf_mutex); |
2364 | return res; | |
171936dd | 2365 | unlock_page: |
761ad8d7 | 2366 | unlock_page(p); |
171936dd TL |
2367 | unlock_mutex: |
2368 | mutex_unlock(&mf_mutex); | |
6a46079c AK |
2369 | return res; |
2370 | } | |
cd42f4a3 | 2371 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 2372 | |
ea8f5fb8 HY |
2373 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
2374 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
2375 | ||
2376 | struct memory_failure_entry { | |
2377 | unsigned long pfn; | |
ea8f5fb8 HY |
2378 | int flags; |
2379 | }; | |
2380 | ||
2381 | struct memory_failure_cpu { | |
2382 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
2383 | MEMORY_FAILURE_FIFO_SIZE); | |
2384 | spinlock_t lock; | |
2385 | struct work_struct work; | |
2386 | }; | |
2387 | ||
2388 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
2389 | ||
2390 | /** | |
2391 | * memory_failure_queue - Schedule handling memory failure of a page. | |
2392 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
2393 | * @flags: Flags for memory failure handling |
2394 | * | |
2395 | * This function is called by the low level hardware error handler | |
2396 | * when it detects hardware memory corruption of a page. It schedules | |
2397 | * the recovering of error page, including dropping pages, killing | |
2398 | * processes etc. | |
2399 | * | |
2400 | * The function is primarily of use for corruptions that | |
2401 | * happen outside the current execution context (e.g. when | |
2402 | * detected by a background scrubber) | |
2403 | * | |
2404 | * Can run in IRQ context. | |
2405 | */ | |
83b57531 | 2406 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
2407 | { |
2408 | struct memory_failure_cpu *mf_cpu; | |
2409 | unsigned long proc_flags; | |
2410 | struct memory_failure_entry entry = { | |
2411 | .pfn = pfn, | |
ea8f5fb8 HY |
2412 | .flags = flags, |
2413 | }; | |
2414 | ||
2415 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
2416 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 2417 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
2418 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
2419 | else | |
96f96763 | 2420 | pr_err("buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
2421 | pfn); |
2422 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
2423 | put_cpu_var(memory_failure_cpu); | |
2424 | } | |
2425 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
2426 | ||
2427 | static void memory_failure_work_func(struct work_struct *work) | |
2428 | { | |
2429 | struct memory_failure_cpu *mf_cpu; | |
2430 | struct memory_failure_entry entry = { 0, }; | |
2431 | unsigned long proc_flags; | |
2432 | int gotten; | |
2433 | ||
06202231 | 2434 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
ea8f5fb8 HY |
2435 | for (;;) { |
2436 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
2437 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
2438 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
2439 | if (!gotten) | |
2440 | break; | |
cf870c70 | 2441 | if (entry.flags & MF_SOFT_OFFLINE) |
feec24a6 | 2442 | soft_offline_page(entry.pfn, entry.flags); |
cf870c70 | 2443 | else |
83b57531 | 2444 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
2445 | } |
2446 | } | |
2447 | ||
06202231 JM |
2448 | /* |
2449 | * Process memory_failure work queued on the specified CPU. | |
2450 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. | |
2451 | */ | |
2452 | void memory_failure_queue_kick(int cpu) | |
2453 | { | |
2454 | struct memory_failure_cpu *mf_cpu; | |
2455 | ||
2456 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
2457 | cancel_work_sync(&mf_cpu->work); | |
2458 | memory_failure_work_func(&mf_cpu->work); | |
2459 | } | |
2460 | ||
ea8f5fb8 HY |
2461 | static int __init memory_failure_init(void) |
2462 | { | |
2463 | struct memory_failure_cpu *mf_cpu; | |
2464 | int cpu; | |
2465 | ||
2466 | for_each_possible_cpu(cpu) { | |
2467 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
2468 | spin_lock_init(&mf_cpu->lock); | |
2469 | INIT_KFIFO(mf_cpu->fifo); | |
2470 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
2471 | } | |
2472 | ||
97de10a9 KW |
2473 | register_sysctl_init("vm", memory_failure_table); |
2474 | ||
ea8f5fb8 HY |
2475 | return 0; |
2476 | } | |
2477 | core_initcall(memory_failure_init); | |
2478 | ||
96f96763 KW |
2479 | #undef pr_fmt |
2480 | #define pr_fmt(fmt) "" fmt | |
a5f65109 NH |
2481 | #define unpoison_pr_info(fmt, pfn, rs) \ |
2482 | ({ \ | |
2483 | if (__ratelimit(rs)) \ | |
2484 | pr_info(fmt, pfn); \ | |
2485 | }) | |
2486 | ||
847ce401 WF |
2487 | /** |
2488 | * unpoison_memory - Unpoison a previously poisoned page | |
2489 | * @pfn: Page number of the to be unpoisoned page | |
2490 | * | |
2491 | * Software-unpoison a page that has been poisoned by | |
2492 | * memory_failure() earlier. | |
2493 | * | |
2494 | * This is only done on the software-level, so it only works | |
2495 | * for linux injected failures, not real hardware failures | |
2496 | * | |
2497 | * Returns 0 for success, otherwise -errno. | |
2498 | */ | |
2499 | int unpoison_memory(unsigned long pfn) | |
2500 | { | |
9637d7df | 2501 | struct folio *folio; |
847ce401 | 2502 | struct page *p; |
f29623e4 | 2503 | int ret = -EBUSY, ghp; |
ac5fcde0 | 2504 | unsigned long count = 1; |
5033091d | 2505 | bool huge = false; |
a5f65109 NH |
2506 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
2507 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
2508 | |
2509 | if (!pfn_valid(pfn)) | |
2510 | return -ENXIO; | |
2511 | ||
2512 | p = pfn_to_page(pfn); | |
9637d7df | 2513 | folio = page_folio(p); |
847ce401 | 2514 | |
91d00547 NH |
2515 | mutex_lock(&mf_mutex); |
2516 | ||
67f22ba7 | 2517 | if (hw_memory_failure) { |
2518 | unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n", | |
2519 | pfn, &unpoison_rs); | |
2520 | ret = -EOPNOTSUPP; | |
2521 | goto unlock_mutex; | |
2522 | } | |
2523 | ||
6c54312f | 2524 | if (!PageHWPoison(p)) { |
495367c0 | 2525 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 2526 | pfn, &unpoison_rs); |
91d00547 | 2527 | goto unlock_mutex; |
847ce401 WF |
2528 | } |
2529 | ||
a6fddef4 | 2530 | if (folio_ref_count(folio) > 1) { |
495367c0 | 2531 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 2532 | pfn, &unpoison_rs); |
91d00547 | 2533 | goto unlock_mutex; |
230ac719 NH |
2534 | } |
2535 | ||
7a8817f2 ML |
2536 | if (folio_test_slab(folio) || PageTable(&folio->page) || |
2537 | folio_test_reserved(folio) || PageOffline(&folio->page)) | |
faeb2ff2 ML |
2538 | goto unlock_mutex; |
2539 | ||
2540 | /* | |
2541 | * Note that folio->_mapcount is overloaded in SLAB, so the simple test | |
2542 | * in folio_mapped() has to be done after folio_test_slab() is checked. | |
2543 | */ | |
a6fddef4 | 2544 | if (folio_mapped(folio)) { |
495367c0 | 2545 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 2546 | pfn, &unpoison_rs); |
91d00547 | 2547 | goto unlock_mutex; |
230ac719 NH |
2548 | } |
2549 | ||
a6fddef4 | 2550 | if (folio_mapping(folio)) { |
495367c0 | 2551 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 2552 | pfn, &unpoison_rs); |
91d00547 | 2553 | goto unlock_mutex; |
0cea3fdc WL |
2554 | } |
2555 | ||
f29623e4 ML |
2556 | ghp = get_hwpoison_page(p, MF_UNPOISON); |
2557 | if (!ghp) { | |
ac5fcde0 | 2558 | if (PageHuge(p)) { |
5033091d | 2559 | huge = true; |
9637d7df | 2560 | count = folio_free_raw_hwp(folio, false); |
f29623e4 | 2561 | if (count == 0) |
ac5fcde0 | 2562 | goto unlock_mutex; |
ac5fcde0 | 2563 | } |
a6fddef4 | 2564 | ret = folio_test_clear_hwpoison(folio) ? 0 : -EBUSY; |
f29623e4 ML |
2565 | } else if (ghp < 0) { |
2566 | if (ghp == -EHWPOISON) { | |
c8bd84f7 | 2567 | ret = put_page_back_buddy(p) ? 0 : -EBUSY; |
f29623e4 ML |
2568 | } else { |
2569 | ret = ghp; | |
bf181c58 NH |
2570 | unpoison_pr_info("Unpoison: failed to grab page %#lx\n", |
2571 | pfn, &unpoison_rs); | |
f29623e4 | 2572 | } |
bf181c58 | 2573 | } else { |
ac5fcde0 | 2574 | if (PageHuge(p)) { |
5033091d | 2575 | huge = true; |
9637d7df | 2576 | count = folio_free_raw_hwp(folio, false); |
ac5fcde0 | 2577 | if (count == 0) { |
a6fddef4 | 2578 | folio_put(folio); |
ac5fcde0 NH |
2579 | goto unlock_mutex; |
2580 | } | |
2581 | } | |
847ce401 | 2582 | |
a6fddef4 | 2583 | folio_put(folio); |
e0ff4280 | 2584 | if (TestClearPageHWPoison(p)) { |
a6fddef4 | 2585 | folio_put(folio); |
bf181c58 NH |
2586 | ret = 0; |
2587 | } | |
2588 | } | |
847ce401 | 2589 | |
91d00547 NH |
2590 | unlock_mutex: |
2591 | mutex_unlock(&mf_mutex); | |
e0ff4280 | 2592 | if (!ret) { |
5033091d NH |
2593 | if (!huge) |
2594 | num_poisoned_pages_sub(pfn, 1); | |
c8bd84f7 | 2595 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", |
2596 | page_to_pfn(p), &unpoison_rs); | |
2597 | } | |
91d00547 | 2598 | return ret; |
847ce401 WF |
2599 | } |
2600 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 2601 | |
6b9a217e | 2602 | static bool isolate_page(struct page *page, struct list_head *pagelist) |
d950b958 | 2603 | { |
6b9a217e | 2604 | bool isolated = false; |
d950b958 | 2605 | |
6b9a217e | 2606 | if (PageHuge(page)) { |
9747b9e9 | 2607 | isolated = isolate_hugetlb(page_folio(page), pagelist); |
6b9a217e | 2608 | } else { |
da294991 ML |
2609 | bool lru = !__PageMovable(page); |
2610 | ||
6b9a217e | 2611 | if (lru) |
f7f9c00d | 2612 | isolated = isolate_lru_page(page); |
6b9a217e | 2613 | else |
cd775580 BW |
2614 | isolated = isolate_movable_page(page, |
2615 | ISOLATE_UNEVICTABLE); | |
6b9a217e | 2616 | |
da294991 | 2617 | if (isolated) { |
6b9a217e | 2618 | list_add(&page->lru, pagelist); |
da294991 ML |
2619 | if (lru) |
2620 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
2621 | page_is_file_lru(page)); | |
2622 | } | |
0ebff32c | 2623 | } |
d950b958 | 2624 | |
03613808 | 2625 | /* |
6b9a217e | 2626 | * If we succeed to isolate the page, we grabbed another refcount on |
5885c6a6 | 2627 | * the page, so we can safely drop the one we got from get_any_page(). |
6b9a217e OS |
2628 | * If we failed to isolate the page, it means that we cannot go further |
2629 | * and we will return an error, so drop the reference we got from | |
5885c6a6 | 2630 | * get_any_page() as well. |
03613808 | 2631 | */ |
6b9a217e OS |
2632 | put_page(page); |
2633 | return isolated; | |
d950b958 NH |
2634 | } |
2635 | ||
6b9a217e | 2636 | /* |
48309e1f | 2637 | * soft_offline_in_use_page handles hugetlb-pages and non-hugetlb pages. |
6b9a217e OS |
2638 | * If the page is a non-dirty unmapped page-cache page, it simply invalidates. |
2639 | * If the page is mapped, it migrates the contents over. | |
2640 | */ | |
48309e1f | 2641 | static int soft_offline_in_use_page(struct page *page) |
af8fae7c | 2642 | { |
d6c75dc2 | 2643 | long ret = 0; |
af8fae7c | 2644 | unsigned long pfn = page_to_pfn(page); |
6b9a217e OS |
2645 | struct page *hpage = compound_head(page); |
2646 | char const *msg_page[] = {"page", "hugepage"}; | |
2647 | bool huge = PageHuge(page); | |
2648 | LIST_HEAD(pagelist); | |
54608759 JK |
2649 | struct migration_target_control mtc = { |
2650 | .nid = NUMA_NO_NODE, | |
2651 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
2652 | }; | |
facb6011 | 2653 | |
48309e1f KW |
2654 | if (!huge && PageTransHuge(hpage)) { |
2655 | if (try_to_split_thp_page(page)) { | |
2656 | pr_info("soft offline: %#lx: thp split failed\n", pfn); | |
2657 | return -EBUSY; | |
2658 | } | |
2659 | hpage = page; | |
2660 | } | |
2661 | ||
0ebff32c | 2662 | lock_page(page); |
55c7ac45 | 2663 | if (!huge) |
6b9a217e | 2664 | wait_on_page_writeback(page); |
af8fae7c NH |
2665 | if (PageHWPoison(page)) { |
2666 | unlock_page(page); | |
dd6e2402 | 2667 | put_page(page); |
af8fae7c | 2668 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
5a2ffca3 | 2669 | return 0; |
af8fae7c | 2670 | } |
6b9a217e | 2671 | |
55c7ac45 | 2672 | if (!huge && PageLRU(page) && !PageSwapCache(page)) |
6b9a217e OS |
2673 | /* |
2674 | * Try to invalidate first. This should work for | |
2675 | * non dirty unmapped page cache pages. | |
2676 | */ | |
2677 | ret = invalidate_inode_page(page); | |
facb6011 | 2678 | unlock_page(page); |
6b9a217e | 2679 | |
6b9a217e | 2680 | if (ret) { |
fb46e735 | 2681 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
6b9a217e | 2682 | page_handle_poison(page, false, true); |
af8fae7c | 2683 | return 0; |
facb6011 AK |
2684 | } |
2685 | ||
6b9a217e | 2686 | if (isolate_page(hpage, &pagelist)) { |
54608759 | 2687 | ret = migrate_pages(&pagelist, alloc_migration_target, NULL, |
5ac95884 | 2688 | (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL); |
79f5f8fa | 2689 | if (!ret) { |
6b9a217e OS |
2690 | bool release = !huge; |
2691 | ||
2692 | if (!page_handle_poison(page, huge, release)) | |
2693 | ret = -EBUSY; | |
79f5f8fa | 2694 | } else { |
85fbe5d1 YX |
2695 | if (!list_empty(&pagelist)) |
2696 | putback_movable_pages(&pagelist); | |
59c82b70 | 2697 | |
d6c75dc2 | 2698 | pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n", |
23efd080 | 2699 | pfn, msg_page[huge], ret, &page->flags); |
facb6011 | 2700 | if (ret > 0) |
3f4b815a | 2701 | ret = -EBUSY; |
facb6011 AK |
2702 | } |
2703 | } else { | |
23efd080 MWO |
2704 | pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n", |
2705 | pfn, msg_page[huge], page_count(page), &page->flags); | |
6b9a217e | 2706 | ret = -EBUSY; |
facb6011 | 2707 | } |
facb6011 AK |
2708 | return ret; |
2709 | } | |
86e05773 WL |
2710 | |
2711 | /** | |
2712 | * soft_offline_page - Soft offline a page. | |
feec24a6 | 2713 | * @pfn: pfn to soft-offline |
86e05773 WL |
2714 | * @flags: flags. Same as memory_failure(). |
2715 | * | |
9113eaf3 | 2716 | * Returns 0 on success |
2717 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event | |
2718 | * < 0 otherwise negated errno. | |
86e05773 WL |
2719 | * |
2720 | * Soft offline a page, by migration or invalidation, | |
2721 | * without killing anything. This is for the case when | |
2722 | * a page is not corrupted yet (so it's still valid to access), | |
2723 | * but has had a number of corrected errors and is better taken | |
2724 | * out. | |
2725 | * | |
2726 | * The actual policy on when to do that is maintained by | |
2727 | * user space. | |
2728 | * | |
2729 | * This should never impact any application or cause data loss, | |
2730 | * however it might take some time. | |
2731 | * | |
2732 | * This is not a 100% solution for all memory, but tries to be | |
2733 | * ``good enough'' for the majority of memory. | |
2734 | */ | |
feec24a6 | 2735 | int soft_offline_page(unsigned long pfn, int flags) |
86e05773 WL |
2736 | { |
2737 | int ret; | |
b94e0282 | 2738 | bool try_again = true; |
b5f1fc98 | 2739 | struct page *page; |
dad4e5b3 | 2740 | |
183a7c5d KW |
2741 | if (!pfn_valid(pfn)) { |
2742 | WARN_ON_ONCE(flags & MF_COUNT_INCREASED); | |
feec24a6 | 2743 | return -ENXIO; |
183a7c5d | 2744 | } |
dad4e5b3 | 2745 | |
feec24a6 NH |
2746 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ |
2747 | page = pfn_to_online_page(pfn); | |
dad4e5b3 | 2748 | if (!page) { |
b5f1fc98 | 2749 | put_ref_page(pfn, flags); |
86a66810 | 2750 | return -EIO; |
dad4e5b3 | 2751 | } |
86a66810 | 2752 | |
91d00547 NH |
2753 | mutex_lock(&mf_mutex); |
2754 | ||
86e05773 | 2755 | if (PageHWPoison(page)) { |
8295d535 | 2756 | pr_info("%s: %#lx page already poisoned\n", __func__, pfn); |
b5f1fc98 | 2757 | put_ref_page(pfn, flags); |
91d00547 | 2758 | mutex_unlock(&mf_mutex); |
5a2ffca3 | 2759 | return 0; |
86e05773 | 2760 | } |
86e05773 | 2761 | |
b94e0282 | 2762 | retry: |
bfc8c901 | 2763 | get_online_mems(); |
bf6445bc | 2764 | ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE); |
bfc8c901 | 2765 | put_online_mems(); |
4e41a30c | 2766 | |
9113eaf3 | 2767 | if (hwpoison_filter(page)) { |
2768 | if (ret > 0) | |
2769 | put_page(page); | |
9113eaf3 | 2770 | |
2771 | mutex_unlock(&mf_mutex); | |
2772 | return -EOPNOTSUPP; | |
2773 | } | |
2774 | ||
8295d535 | 2775 | if (ret > 0) { |
6b9a217e | 2776 | ret = soft_offline_in_use_page(page); |
8295d535 | 2777 | } else if (ret == 0) { |
e2c1ab07 ML |
2778 | if (!page_handle_poison(page, true, false)) { |
2779 | if (try_again) { | |
2780 | try_again = false; | |
2781 | flags &= ~MF_COUNT_INCREASED; | |
2782 | goto retry; | |
2783 | } | |
2784 | ret = -EBUSY; | |
b94e0282 | 2785 | } |
8295d535 | 2786 | } |
4e41a30c | 2787 | |
91d00547 NH |
2788 | mutex_unlock(&mf_mutex); |
2789 | ||
86e05773 WL |
2790 | return ret; |
2791 | } |