Commit | Line | Data |
---|---|---|
1439f94c | 1 | // SPDX-License-Identifier: GPL-2.0-only |
6a46079c AK |
2 | /* |
3 | * Copyright (C) 2008, 2009 Intel Corporation | |
4 | * Authors: Andi Kleen, Fengguang Wu | |
5 | * | |
6a46079c | 6 | * High level machine check handler. Handles pages reported by the |
1c80b990 | 7 | * hardware as being corrupted usually due to a multi-bit ECC memory or cache |
6a46079c | 8 | * failure. |
1c80b990 AK |
9 | * |
10 | * In addition there is a "soft offline" entry point that allows stop using | |
11 | * not-yet-corrupted-by-suspicious pages without killing anything. | |
6a46079c AK |
12 | * |
13 | * Handles page cache pages in various states. The tricky part | |
1c80b990 AK |
14 | * here is that we can access any page asynchronously in respect to |
15 | * other VM users, because memory failures could happen anytime and | |
16 | * anywhere. This could violate some of their assumptions. This is why | |
17 | * this code has to be extremely careful. Generally it tries to use | |
18 | * normal locking rules, as in get the standard locks, even if that means | |
19 | * the error handling takes potentially a long time. | |
e0de78df AK |
20 | * |
21 | * It can be very tempting to add handling for obscure cases here. | |
22 | * In general any code for handling new cases should only be added iff: | |
23 | * - You know how to test it. | |
24 | * - You have a test that can be added to mce-test | |
25 | * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/ | |
26 | * - The case actually shows up as a frequent (top 10) page state in | |
27 | * tools/vm/page-types when running a real workload. | |
1c80b990 AK |
28 | * |
29 | * There are several operations here with exponential complexity because | |
30 | * of unsuitable VM data structures. For example the operation to map back | |
31 | * from RMAP chains to processes has to walk the complete process list and | |
32 | * has non linear complexity with the number. But since memory corruptions | |
33 | * are rare we hope to get away with this. This avoids impacting the core | |
34 | * VM. | |
6a46079c | 35 | */ |
6a46079c AK |
36 | #include <linux/kernel.h> |
37 | #include <linux/mm.h> | |
38 | #include <linux/page-flags.h> | |
478c5ffc | 39 | #include <linux/kernel-page-flags.h> |
3f07c014 | 40 | #include <linux/sched/signal.h> |
29930025 | 41 | #include <linux/sched/task.h> |
01e00f88 | 42 | #include <linux/ksm.h> |
6a46079c | 43 | #include <linux/rmap.h> |
b9e15baf | 44 | #include <linux/export.h> |
6a46079c AK |
45 | #include <linux/pagemap.h> |
46 | #include <linux/swap.h> | |
47 | #include <linux/backing-dev.h> | |
facb6011 | 48 | #include <linux/migrate.h> |
facb6011 | 49 | #include <linux/suspend.h> |
5a0e3ad6 | 50 | #include <linux/slab.h> |
bf998156 | 51 | #include <linux/swapops.h> |
7af446a8 | 52 | #include <linux/hugetlb.h> |
20d6c96b | 53 | #include <linux/memory_hotplug.h> |
5db8a73a | 54 | #include <linux/mm_inline.h> |
6100e34b | 55 | #include <linux/memremap.h> |
ea8f5fb8 | 56 | #include <linux/kfifo.h> |
a5f65109 | 57 | #include <linux/ratelimit.h> |
d4ae9916 | 58 | #include <linux/page-isolation.h> |
6a46079c | 59 | #include "internal.h" |
97f0b134 | 60 | #include "ras/ras_event.h" |
6a46079c AK |
61 | |
62 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
63 | ||
64 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
65 | ||
293c07e3 | 66 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 67 | |
6b9a217e | 68 | static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release) |
06be6ff3 | 69 | { |
6b9a217e OS |
70 | if (hugepage_or_freepage) { |
71 | /* | |
72 | * Doing this check for free pages is also fine since dissolve_free_huge_page | |
73 | * returns 0 for non-hugetlb pages as well. | |
74 | */ | |
75 | if (dissolve_free_huge_page(page) || !take_page_off_buddy(page)) | |
76 | /* | |
77 | * We could fail to take off the target page from buddy | |
78 | * for example due to racy page allocaiton, but that's | |
79 | * acceptable because soft-offlined page is not broken | |
80 | * and if someone really want to use it, they should | |
81 | * take it. | |
82 | */ | |
83 | return false; | |
84 | } | |
85 | ||
06be6ff3 | 86 | SetPageHWPoison(page); |
79f5f8fa OS |
87 | if (release) |
88 | put_page(page); | |
06be6ff3 OS |
89 | page_ref_inc(page); |
90 | num_poisoned_pages_inc(); | |
6b9a217e OS |
91 | |
92 | return true; | |
06be6ff3 OS |
93 | } |
94 | ||
27df5068 AK |
95 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
96 | ||
1bfe5feb | 97 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
98 | u32 hwpoison_filter_dev_major = ~0U; |
99 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
100 | u64 hwpoison_filter_flags_mask; |
101 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 102 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
103 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
104 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
105 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
106 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
107 | |
108 | static int hwpoison_filter_dev(struct page *p) | |
109 | { | |
110 | struct address_space *mapping; | |
111 | dev_t dev; | |
112 | ||
113 | if (hwpoison_filter_dev_major == ~0U && | |
114 | hwpoison_filter_dev_minor == ~0U) | |
115 | return 0; | |
116 | ||
117 | /* | |
1c80b990 | 118 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
119 | */ |
120 | if (PageSlab(p)) | |
121 | return -EINVAL; | |
122 | ||
123 | mapping = page_mapping(p); | |
124 | if (mapping == NULL || mapping->host == NULL) | |
125 | return -EINVAL; | |
126 | ||
127 | dev = mapping->host->i_sb->s_dev; | |
128 | if (hwpoison_filter_dev_major != ~0U && | |
129 | hwpoison_filter_dev_major != MAJOR(dev)) | |
130 | return -EINVAL; | |
131 | if (hwpoison_filter_dev_minor != ~0U && | |
132 | hwpoison_filter_dev_minor != MINOR(dev)) | |
133 | return -EINVAL; | |
134 | ||
135 | return 0; | |
136 | } | |
137 | ||
478c5ffc WF |
138 | static int hwpoison_filter_flags(struct page *p) |
139 | { | |
140 | if (!hwpoison_filter_flags_mask) | |
141 | return 0; | |
142 | ||
143 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
144 | hwpoison_filter_flags_value) | |
145 | return 0; | |
146 | else | |
147 | return -EINVAL; | |
148 | } | |
149 | ||
4fd466eb AK |
150 | /* |
151 | * This allows stress tests to limit test scope to a collection of tasks | |
152 | * by putting them under some memcg. This prevents killing unrelated/important | |
153 | * processes such as /sbin/init. Note that the target task may share clean | |
154 | * pages with init (eg. libc text), which is harmless. If the target task | |
155 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
156 | * is also included in the memcg. At last, due to race conditions this filter | |
157 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
158 | * a freed page. | |
159 | */ | |
94a59fb3 | 160 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
161 | u64 hwpoison_filter_memcg; |
162 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
163 | static int hwpoison_filter_task(struct page *p) | |
164 | { | |
4fd466eb AK |
165 | if (!hwpoison_filter_memcg) |
166 | return 0; | |
167 | ||
94a59fb3 | 168 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
169 | return -EINVAL; |
170 | ||
171 | return 0; | |
172 | } | |
173 | #else | |
174 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
175 | #endif | |
176 | ||
7c116f2b WF |
177 | int hwpoison_filter(struct page *p) |
178 | { | |
1bfe5feb HL |
179 | if (!hwpoison_filter_enable) |
180 | return 0; | |
181 | ||
7c116f2b WF |
182 | if (hwpoison_filter_dev(p)) |
183 | return -EINVAL; | |
184 | ||
478c5ffc WF |
185 | if (hwpoison_filter_flags(p)) |
186 | return -EINVAL; | |
187 | ||
4fd466eb AK |
188 | if (hwpoison_filter_task(p)) |
189 | return -EINVAL; | |
190 | ||
7c116f2b WF |
191 | return 0; |
192 | } | |
27df5068 AK |
193 | #else |
194 | int hwpoison_filter(struct page *p) | |
195 | { | |
196 | return 0; | |
197 | } | |
198 | #endif | |
199 | ||
7c116f2b WF |
200 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
201 | ||
ae1139ec DW |
202 | /* |
203 | * Kill all processes that have a poisoned page mapped and then isolate | |
204 | * the page. | |
205 | * | |
206 | * General strategy: | |
207 | * Find all processes having the page mapped and kill them. | |
208 | * But we keep a page reference around so that the page is not | |
209 | * actually freed yet. | |
210 | * Then stash the page away | |
211 | * | |
212 | * There's no convenient way to get back to mapped processes | |
213 | * from the VMAs. So do a brute-force search over all | |
214 | * running processes. | |
215 | * | |
216 | * Remember that machine checks are not common (or rather | |
217 | * if they are common you have other problems), so this shouldn't | |
218 | * be a performance issue. | |
219 | * | |
220 | * Also there are some races possible while we get from the | |
221 | * error detection to actually handle it. | |
222 | */ | |
223 | ||
224 | struct to_kill { | |
225 | struct list_head nd; | |
226 | struct task_struct *tsk; | |
227 | unsigned long addr; | |
228 | short size_shift; | |
ae1139ec DW |
229 | }; |
230 | ||
6a46079c | 231 | /* |
7329bbeb TL |
232 | * Send all the processes who have the page mapped a signal. |
233 | * ``action optional'' if they are not immediately affected by the error | |
234 | * ``action required'' if error happened in current execution context | |
6a46079c | 235 | */ |
ae1139ec | 236 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 237 | { |
ae1139ec DW |
238 | struct task_struct *t = tk->tsk; |
239 | short addr_lsb = tk->size_shift; | |
872e9a20 | 240 | int ret = 0; |
6a46079c | 241 | |
03151c6e | 242 | pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", |
872e9a20 | 243 | pfn, t->comm, t->pid); |
7329bbeb | 244 | |
872e9a20 | 245 | if (flags & MF_ACTION_REQUIRED) { |
03151c6e NH |
246 | WARN_ON_ONCE(t != current); |
247 | ret = force_sig_mceerr(BUS_MCEERR_AR, | |
872e9a20 | 248 | (void __user *)tk->addr, addr_lsb); |
7329bbeb TL |
249 | } else { |
250 | /* | |
251 | * Don't use force here, it's convenient if the signal | |
252 | * can be temporarily blocked. | |
253 | * This could cause a loop when the user sets SIGBUS | |
254 | * to SIG_IGN, but hopefully no one will do that? | |
255 | */ | |
ae1139ec | 256 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
c0f45555 | 257 | addr_lsb, t); /* synchronous? */ |
7329bbeb | 258 | } |
6a46079c | 259 | if (ret < 0) |
495367c0 | 260 | pr_info("Memory failure: Error sending signal to %s:%d: %d\n", |
1170532b | 261 | t->comm, t->pid, ret); |
6a46079c AK |
262 | return ret; |
263 | } | |
264 | ||
588f9ce6 | 265 | /* |
47e431f4 OS |
266 | * Unknown page type encountered. Try to check whether it can turn PageLRU by |
267 | * lru_add_drain_all, or a free page by reclaiming slabs when possible. | |
588f9ce6 | 268 | */ |
facb6011 | 269 | void shake_page(struct page *p, int access) |
588f9ce6 | 270 | { |
8bcb74de NH |
271 | if (PageHuge(p)) |
272 | return; | |
273 | ||
588f9ce6 AK |
274 | if (!PageSlab(p)) { |
275 | lru_add_drain_all(); | |
588f9ce6 AK |
276 | if (PageLRU(p) || is_free_buddy_page(p)) |
277 | return; | |
278 | } | |
facb6011 | 279 | |
588f9ce6 | 280 | /* |
6b4f7799 JW |
281 | * Only call shrink_node_slabs here (which would also shrink |
282 | * other caches) if access is not potentially fatal. | |
588f9ce6 | 283 | */ |
cb731d6c VD |
284 | if (access) |
285 | drop_slab_node(page_to_nid(p)); | |
588f9ce6 AK |
286 | } |
287 | EXPORT_SYMBOL_GPL(shake_page); | |
288 | ||
6100e34b DW |
289 | static unsigned long dev_pagemap_mapping_shift(struct page *page, |
290 | struct vm_area_struct *vma) | |
291 | { | |
292 | unsigned long address = vma_address(page, vma); | |
293 | pgd_t *pgd; | |
294 | p4d_t *p4d; | |
295 | pud_t *pud; | |
296 | pmd_t *pmd; | |
297 | pte_t *pte; | |
298 | ||
299 | pgd = pgd_offset(vma->vm_mm, address); | |
300 | if (!pgd_present(*pgd)) | |
301 | return 0; | |
302 | p4d = p4d_offset(pgd, address); | |
303 | if (!p4d_present(*p4d)) | |
304 | return 0; | |
305 | pud = pud_offset(p4d, address); | |
306 | if (!pud_present(*pud)) | |
307 | return 0; | |
308 | if (pud_devmap(*pud)) | |
309 | return PUD_SHIFT; | |
310 | pmd = pmd_offset(pud, address); | |
311 | if (!pmd_present(*pmd)) | |
312 | return 0; | |
313 | if (pmd_devmap(*pmd)) | |
314 | return PMD_SHIFT; | |
315 | pte = pte_offset_map(pmd, address); | |
316 | if (!pte_present(*pte)) | |
317 | return 0; | |
318 | if (pte_devmap(*pte)) | |
319 | return PAGE_SHIFT; | |
320 | return 0; | |
321 | } | |
6a46079c AK |
322 | |
323 | /* | |
324 | * Failure handling: if we can't find or can't kill a process there's | |
325 | * not much we can do. We just print a message and ignore otherwise. | |
326 | */ | |
327 | ||
328 | /* | |
329 | * Schedule a process for later kill. | |
330 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
6a46079c AK |
331 | */ |
332 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
333 | struct vm_area_struct *vma, | |
996ff7a0 | 334 | struct list_head *to_kill) |
6a46079c AK |
335 | { |
336 | struct to_kill *tk; | |
337 | ||
996ff7a0 JC |
338 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
339 | if (!tk) { | |
340 | pr_err("Memory failure: Out of memory while machine check handling\n"); | |
341 | return; | |
6a46079c | 342 | } |
996ff7a0 | 343 | |
6a46079c | 344 | tk->addr = page_address_in_vma(p, vma); |
6100e34b DW |
345 | if (is_zone_device_page(p)) |
346 | tk->size_shift = dev_pagemap_mapping_shift(p, vma); | |
347 | else | |
75068518 | 348 | tk->size_shift = page_shift(compound_head(p)); |
6a46079c AK |
349 | |
350 | /* | |
3d7fed4a JC |
351 | * Send SIGKILL if "tk->addr == -EFAULT". Also, as |
352 | * "tk->size_shift" is always non-zero for !is_zone_device_page(), | |
353 | * so "tk->size_shift == 0" effectively checks no mapping on | |
354 | * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times | |
355 | * to a process' address space, it's possible not all N VMAs | |
356 | * contain mappings for the page, but at least one VMA does. | |
357 | * Only deliver SIGBUS with payload derived from the VMA that | |
358 | * has a mapping for the page. | |
6a46079c | 359 | */ |
3d7fed4a | 360 | if (tk->addr == -EFAULT) { |
495367c0 | 361 | pr_info("Memory failure: Unable to find user space address %lx in %s\n", |
6a46079c | 362 | page_to_pfn(p), tsk->comm); |
3d7fed4a JC |
363 | } else if (tk->size_shift == 0) { |
364 | kfree(tk); | |
365 | return; | |
6a46079c | 366 | } |
996ff7a0 | 367 | |
6a46079c AK |
368 | get_task_struct(tsk); |
369 | tk->tsk = tsk; | |
370 | list_add_tail(&tk->nd, to_kill); | |
371 | } | |
372 | ||
373 | /* | |
374 | * Kill the processes that have been collected earlier. | |
375 | * | |
376 | * Only do anything when DOIT is set, otherwise just free the list | |
377 | * (this is used for clean pages which do not need killing) | |
378 | * Also when FAIL is set do a force kill because something went | |
379 | * wrong earlier. | |
380 | */ | |
ae1139ec DW |
381 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
382 | unsigned long pfn, int flags) | |
6a46079c AK |
383 | { |
384 | struct to_kill *tk, *next; | |
385 | ||
386 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 387 | if (forcekill) { |
6a46079c | 388 | /* |
af901ca1 | 389 | * In case something went wrong with munmapping |
6a46079c AK |
390 | * make sure the process doesn't catch the |
391 | * signal and then access the memory. Just kill it. | |
6a46079c | 392 | */ |
3d7fed4a | 393 | if (fail || tk->addr == -EFAULT) { |
495367c0 | 394 | pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 395 | pfn, tk->tsk->comm, tk->tsk->pid); |
6376360e NH |
396 | do_send_sig_info(SIGKILL, SEND_SIG_PRIV, |
397 | tk->tsk, PIDTYPE_PID); | |
6a46079c AK |
398 | } |
399 | ||
400 | /* | |
401 | * In theory the process could have mapped | |
402 | * something else on the address in-between. We could | |
403 | * check for that, but we need to tell the | |
404 | * process anyways. | |
405 | */ | |
ae1139ec | 406 | else if (kill_proc(tk, pfn, flags) < 0) |
495367c0 | 407 | pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 408 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c AK |
409 | } |
410 | put_task_struct(tk->tsk); | |
411 | kfree(tk); | |
412 | } | |
413 | } | |
414 | ||
3ba08129 NH |
415 | /* |
416 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
417 | * on behalf of the thread group. Return task_struct of the (first found) | |
418 | * dedicated thread if found, and return NULL otherwise. | |
419 | * | |
420 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
421 | * have to call rcu_read_lock/unlock() in this function. | |
422 | */ | |
423 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 424 | { |
3ba08129 NH |
425 | struct task_struct *t; |
426 | ||
4e018b45 NH |
427 | for_each_thread(tsk, t) { |
428 | if (t->flags & PF_MCE_PROCESS) { | |
429 | if (t->flags & PF_MCE_EARLY) | |
430 | return t; | |
431 | } else { | |
432 | if (sysctl_memory_failure_early_kill) | |
433 | return t; | |
434 | } | |
435 | } | |
3ba08129 NH |
436 | return NULL; |
437 | } | |
438 | ||
439 | /* | |
440 | * Determine whether a given process is "early kill" process which expects | |
441 | * to be signaled when some page under the process is hwpoisoned. | |
442 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
443 | * specified) if the process is "early kill," and otherwise returns NULL. | |
03151c6e NH |
444 | * |
445 | * Note that the above is true for Action Optional case, but not for Action | |
446 | * Required case where SIGBUS should sent only to the current thread. | |
3ba08129 NH |
447 | */ |
448 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
449 | int force_early) | |
450 | { | |
6a46079c | 451 | if (!tsk->mm) |
3ba08129 | 452 | return NULL; |
03151c6e NH |
453 | if (force_early) { |
454 | /* | |
455 | * Comparing ->mm here because current task might represent | |
456 | * a subthread, while tsk always points to the main thread. | |
457 | */ | |
458 | if (tsk->mm == current->mm) | |
459 | return current; | |
460 | else | |
461 | return NULL; | |
462 | } | |
4e018b45 | 463 | return find_early_kill_thread(tsk); |
6a46079c AK |
464 | } |
465 | ||
466 | /* | |
467 | * Collect processes when the error hit an anonymous page. | |
468 | */ | |
469 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 470 | int force_early) |
6a46079c AK |
471 | { |
472 | struct vm_area_struct *vma; | |
473 | struct task_struct *tsk; | |
474 | struct anon_vma *av; | |
bf181b9f | 475 | pgoff_t pgoff; |
6a46079c | 476 | |
4fc3f1d6 | 477 | av = page_lock_anon_vma_read(page); |
6a46079c | 478 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
479 | return; |
480 | ||
a0f7a756 | 481 | pgoff = page_to_pgoff(page); |
9b679320 | 482 | read_lock(&tasklist_lock); |
6a46079c | 483 | for_each_process (tsk) { |
5beb4930 | 484 | struct anon_vma_chain *vmac; |
3ba08129 | 485 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 486 | |
3ba08129 | 487 | if (!t) |
6a46079c | 488 | continue; |
bf181b9f ML |
489 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
490 | pgoff, pgoff) { | |
5beb4930 | 491 | vma = vmac->vma; |
6a46079c AK |
492 | if (!page_mapped_in_vma(page, vma)) |
493 | continue; | |
3ba08129 | 494 | if (vma->vm_mm == t->mm) |
996ff7a0 | 495 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
496 | } |
497 | } | |
6a46079c | 498 | read_unlock(&tasklist_lock); |
4fc3f1d6 | 499 | page_unlock_anon_vma_read(av); |
6a46079c AK |
500 | } |
501 | ||
502 | /* | |
503 | * Collect processes when the error hit a file mapped page. | |
504 | */ | |
505 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 506 | int force_early) |
6a46079c AK |
507 | { |
508 | struct vm_area_struct *vma; | |
509 | struct task_struct *tsk; | |
6a46079c | 510 | struct address_space *mapping = page->mapping; |
c43bc03d | 511 | pgoff_t pgoff; |
6a46079c | 512 | |
d28eb9c8 | 513 | i_mmap_lock_read(mapping); |
9b679320 | 514 | read_lock(&tasklist_lock); |
c43bc03d | 515 | pgoff = page_to_pgoff(page); |
6a46079c | 516 | for_each_process(tsk) { |
3ba08129 | 517 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 518 | |
3ba08129 | 519 | if (!t) |
6a46079c | 520 | continue; |
6b2dbba8 | 521 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
522 | pgoff) { |
523 | /* | |
524 | * Send early kill signal to tasks where a vma covers | |
525 | * the page but the corrupted page is not necessarily | |
526 | * mapped it in its pte. | |
527 | * Assume applications who requested early kill want | |
528 | * to be informed of all such data corruptions. | |
529 | */ | |
3ba08129 | 530 | if (vma->vm_mm == t->mm) |
996ff7a0 | 531 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
532 | } |
533 | } | |
6a46079c | 534 | read_unlock(&tasklist_lock); |
d28eb9c8 | 535 | i_mmap_unlock_read(mapping); |
6a46079c AK |
536 | } |
537 | ||
538 | /* | |
539 | * Collect the processes who have the corrupted page mapped to kill. | |
6a46079c | 540 | */ |
74614de1 TL |
541 | static void collect_procs(struct page *page, struct list_head *tokill, |
542 | int force_early) | |
6a46079c | 543 | { |
6a46079c AK |
544 | if (!page->mapping) |
545 | return; | |
546 | ||
6a46079c | 547 | if (PageAnon(page)) |
996ff7a0 | 548 | collect_procs_anon(page, tokill, force_early); |
6a46079c | 549 | else |
996ff7a0 | 550 | collect_procs_file(page, tokill, force_early); |
6a46079c AK |
551 | } |
552 | ||
6a46079c | 553 | static const char *action_name[] = { |
cc637b17 XX |
554 | [MF_IGNORED] = "Ignored", |
555 | [MF_FAILED] = "Failed", | |
556 | [MF_DELAYED] = "Delayed", | |
557 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
558 | }; |
559 | ||
560 | static const char * const action_page_types[] = { | |
cc637b17 XX |
561 | [MF_MSG_KERNEL] = "reserved kernel page", |
562 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
563 | [MF_MSG_SLAB] = "kernel slab page", | |
564 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
565 | [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", | |
566 | [MF_MSG_HUGE] = "huge page", | |
567 | [MF_MSG_FREE_HUGE] = "free huge page", | |
31286a84 | 568 | [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page", |
cc637b17 XX |
569 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", |
570 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
571 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
572 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
573 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
574 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
575 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
576 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
577 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
578 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
579 | [MF_MSG_BUDDY] = "free buddy page", | |
580 | [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", | |
6100e34b | 581 | [MF_MSG_DAX] = "dax page", |
5d1fd5dc | 582 | [MF_MSG_UNSPLIT_THP] = "unsplit thp", |
cc637b17 | 583 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
584 | }; |
585 | ||
dc2a1cbf WF |
586 | /* |
587 | * XXX: It is possible that a page is isolated from LRU cache, | |
588 | * and then kept in swap cache or failed to remove from page cache. | |
589 | * The page count will stop it from being freed by unpoison. | |
590 | * Stress tests should be aware of this memory leak problem. | |
591 | */ | |
592 | static int delete_from_lru_cache(struct page *p) | |
593 | { | |
594 | if (!isolate_lru_page(p)) { | |
595 | /* | |
596 | * Clear sensible page flags, so that the buddy system won't | |
597 | * complain when the page is unpoison-and-freed. | |
598 | */ | |
599 | ClearPageActive(p); | |
600 | ClearPageUnevictable(p); | |
18365225 MH |
601 | |
602 | /* | |
603 | * Poisoned page might never drop its ref count to 0 so we have | |
604 | * to uncharge it manually from its memcg. | |
605 | */ | |
606 | mem_cgroup_uncharge(p); | |
607 | ||
dc2a1cbf WF |
608 | /* |
609 | * drop the page count elevated by isolate_lru_page() | |
610 | */ | |
09cbfeaf | 611 | put_page(p); |
dc2a1cbf WF |
612 | return 0; |
613 | } | |
614 | return -EIO; | |
615 | } | |
616 | ||
78bb9203 NH |
617 | static int truncate_error_page(struct page *p, unsigned long pfn, |
618 | struct address_space *mapping) | |
619 | { | |
620 | int ret = MF_FAILED; | |
621 | ||
622 | if (mapping->a_ops->error_remove_page) { | |
623 | int err = mapping->a_ops->error_remove_page(mapping, p); | |
624 | ||
625 | if (err != 0) { | |
626 | pr_info("Memory failure: %#lx: Failed to punch page: %d\n", | |
627 | pfn, err); | |
628 | } else if (page_has_private(p) && | |
629 | !try_to_release_page(p, GFP_NOIO)) { | |
630 | pr_info("Memory failure: %#lx: failed to release buffers\n", | |
631 | pfn); | |
632 | } else { | |
633 | ret = MF_RECOVERED; | |
634 | } | |
635 | } else { | |
636 | /* | |
637 | * If the file system doesn't support it just invalidate | |
638 | * This fails on dirty or anything with private pages | |
639 | */ | |
640 | if (invalidate_inode_page(p)) | |
641 | ret = MF_RECOVERED; | |
642 | else | |
643 | pr_info("Memory failure: %#lx: Failed to invalidate\n", | |
644 | pfn); | |
645 | } | |
646 | ||
647 | return ret; | |
648 | } | |
649 | ||
6a46079c AK |
650 | /* |
651 | * Error hit kernel page. | |
652 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
653 | * could be more sophisticated. | |
654 | */ | |
655 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c | 656 | { |
cc637b17 | 657 | return MF_IGNORED; |
6a46079c AK |
658 | } |
659 | ||
660 | /* | |
661 | * Page in unknown state. Do nothing. | |
662 | */ | |
663 | static int me_unknown(struct page *p, unsigned long pfn) | |
664 | { | |
495367c0 | 665 | pr_err("Memory failure: %#lx: Unknown page state\n", pfn); |
cc637b17 | 666 | return MF_FAILED; |
6a46079c AK |
667 | } |
668 | ||
6a46079c AK |
669 | /* |
670 | * Clean (or cleaned) page cache page. | |
671 | */ | |
672 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
673 | { | |
6a46079c AK |
674 | struct address_space *mapping; |
675 | ||
dc2a1cbf WF |
676 | delete_from_lru_cache(p); |
677 | ||
6a46079c AK |
678 | /* |
679 | * For anonymous pages we're done the only reference left | |
680 | * should be the one m_f() holds. | |
681 | */ | |
682 | if (PageAnon(p)) | |
cc637b17 | 683 | return MF_RECOVERED; |
6a46079c AK |
684 | |
685 | /* | |
686 | * Now truncate the page in the page cache. This is really | |
687 | * more like a "temporary hole punch" | |
688 | * Don't do this for block devices when someone else | |
689 | * has a reference, because it could be file system metadata | |
690 | * and that's not safe to truncate. | |
691 | */ | |
692 | mapping = page_mapping(p); | |
693 | if (!mapping) { | |
694 | /* | |
695 | * Page has been teared down in the meanwhile | |
696 | */ | |
cc637b17 | 697 | return MF_FAILED; |
6a46079c AK |
698 | } |
699 | ||
700 | /* | |
701 | * Truncation is a bit tricky. Enable it per file system for now. | |
702 | * | |
703 | * Open: to take i_mutex or not for this? Right now we don't. | |
704 | */ | |
78bb9203 | 705 | return truncate_error_page(p, pfn, mapping); |
6a46079c AK |
706 | } |
707 | ||
708 | /* | |
549543df | 709 | * Dirty pagecache page |
6a46079c AK |
710 | * Issues: when the error hit a hole page the error is not properly |
711 | * propagated. | |
712 | */ | |
713 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
714 | { | |
715 | struct address_space *mapping = page_mapping(p); | |
716 | ||
717 | SetPageError(p); | |
718 | /* TBD: print more information about the file. */ | |
719 | if (mapping) { | |
720 | /* | |
721 | * IO error will be reported by write(), fsync(), etc. | |
722 | * who check the mapping. | |
723 | * This way the application knows that something went | |
724 | * wrong with its dirty file data. | |
725 | * | |
726 | * There's one open issue: | |
727 | * | |
728 | * The EIO will be only reported on the next IO | |
729 | * operation and then cleared through the IO map. | |
730 | * Normally Linux has two mechanisms to pass IO error | |
731 | * first through the AS_EIO flag in the address space | |
732 | * and then through the PageError flag in the page. | |
733 | * Since we drop pages on memory failure handling the | |
734 | * only mechanism open to use is through AS_AIO. | |
735 | * | |
736 | * This has the disadvantage that it gets cleared on | |
737 | * the first operation that returns an error, while | |
738 | * the PageError bit is more sticky and only cleared | |
739 | * when the page is reread or dropped. If an | |
740 | * application assumes it will always get error on | |
741 | * fsync, but does other operations on the fd before | |
25985edc | 742 | * and the page is dropped between then the error |
6a46079c AK |
743 | * will not be properly reported. |
744 | * | |
745 | * This can already happen even without hwpoisoned | |
746 | * pages: first on metadata IO errors (which only | |
747 | * report through AS_EIO) or when the page is dropped | |
748 | * at the wrong time. | |
749 | * | |
750 | * So right now we assume that the application DTRT on | |
751 | * the first EIO, but we're not worse than other parts | |
752 | * of the kernel. | |
753 | */ | |
af21bfaf | 754 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
755 | } |
756 | ||
757 | return me_pagecache_clean(p, pfn); | |
758 | } | |
759 | ||
760 | /* | |
761 | * Clean and dirty swap cache. | |
762 | * | |
763 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
764 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
765 | * referenced concurrently by 2 types of PTEs: | |
766 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
767 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
768 | * and then | |
769 | * - clear dirty bit to prevent IO | |
770 | * - remove from LRU | |
771 | * - but keep in the swap cache, so that when we return to it on | |
772 | * a later page fault, we know the application is accessing | |
773 | * corrupted data and shall be killed (we installed simple | |
774 | * interception code in do_swap_page to catch it). | |
775 | * | |
776 | * Clean swap cache pages can be directly isolated. A later page fault will | |
777 | * bring in the known good data from disk. | |
778 | */ | |
779 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
780 | { | |
6a46079c AK |
781 | ClearPageDirty(p); |
782 | /* Trigger EIO in shmem: */ | |
783 | ClearPageUptodate(p); | |
784 | ||
dc2a1cbf | 785 | if (!delete_from_lru_cache(p)) |
cc637b17 | 786 | return MF_DELAYED; |
dc2a1cbf | 787 | else |
cc637b17 | 788 | return MF_FAILED; |
6a46079c AK |
789 | } |
790 | ||
791 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
792 | { | |
6a46079c | 793 | delete_from_swap_cache(p); |
e43c3afb | 794 | |
dc2a1cbf | 795 | if (!delete_from_lru_cache(p)) |
cc637b17 | 796 | return MF_RECOVERED; |
dc2a1cbf | 797 | else |
cc637b17 | 798 | return MF_FAILED; |
6a46079c AK |
799 | } |
800 | ||
801 | /* | |
802 | * Huge pages. Needs work. | |
803 | * Issues: | |
93f70f90 NH |
804 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
805 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
806 | */ |
807 | static int me_huge_page(struct page *p, unsigned long pfn) | |
808 | { | |
a8b2c2ce | 809 | int res; |
93f70f90 | 810 | struct page *hpage = compound_head(p); |
78bb9203 | 811 | struct address_space *mapping; |
2491ffee NH |
812 | |
813 | if (!PageHuge(hpage)) | |
814 | return MF_DELAYED; | |
815 | ||
78bb9203 NH |
816 | mapping = page_mapping(hpage); |
817 | if (mapping) { | |
818 | res = truncate_error_page(hpage, pfn, mapping); | |
819 | } else { | |
a8b2c2ce | 820 | res = MF_FAILED; |
78bb9203 NH |
821 | unlock_page(hpage); |
822 | /* | |
823 | * migration entry prevents later access on error anonymous | |
824 | * hugepage, so we can free and dissolve it into buddy to | |
825 | * save healthy subpages. | |
826 | */ | |
827 | if (PageAnon(hpage)) | |
828 | put_page(hpage); | |
a8b2c2ce OS |
829 | if (!dissolve_free_huge_page(p) && take_page_off_buddy(p)) { |
830 | page_ref_inc(p); | |
831 | res = MF_RECOVERED; | |
832 | } | |
78bb9203 | 833 | lock_page(hpage); |
93f70f90 | 834 | } |
78bb9203 NH |
835 | |
836 | return res; | |
6a46079c AK |
837 | } |
838 | ||
839 | /* | |
840 | * Various page states we can handle. | |
841 | * | |
842 | * A page state is defined by its current page->flags bits. | |
843 | * The table matches them in order and calls the right handler. | |
844 | * | |
845 | * This is quite tricky because we can access page at any time | |
25985edc | 846 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
847 | * |
848 | * This is not complete. More states could be added. | |
849 | * For any missing state don't attempt recovery. | |
850 | */ | |
851 | ||
852 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 853 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
854 | #define unevict (1UL << PG_unevictable) |
855 | #define mlock (1UL << PG_mlocked) | |
6a46079c | 856 | #define lru (1UL << PG_lru) |
6a46079c | 857 | #define head (1UL << PG_head) |
6a46079c | 858 | #define slab (1UL << PG_slab) |
6a46079c AK |
859 | #define reserved (1UL << PG_reserved) |
860 | ||
861 | static struct page_state { | |
862 | unsigned long mask; | |
863 | unsigned long res; | |
cc637b17 | 864 | enum mf_action_page_type type; |
6a46079c AK |
865 | int (*action)(struct page *p, unsigned long pfn); |
866 | } error_states[] = { | |
cc637b17 | 867 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
868 | /* |
869 | * free pages are specially detected outside this table: | |
870 | * PG_buddy pages only make a small fraction of all free pages. | |
871 | */ | |
6a46079c AK |
872 | |
873 | /* | |
874 | * Could in theory check if slab page is free or if we can drop | |
875 | * currently unused objects without touching them. But just | |
876 | * treat it as standard kernel for now. | |
877 | */ | |
cc637b17 | 878 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 879 | |
cc637b17 | 880 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 881 | |
cc637b17 XX |
882 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
883 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 884 | |
cc637b17 XX |
885 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
886 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 887 | |
cc637b17 XX |
888 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
889 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 890 | |
cc637b17 XX |
891 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
892 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
893 | |
894 | /* | |
895 | * Catchall entry: must be at end. | |
896 | */ | |
cc637b17 | 897 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
898 | }; |
899 | ||
2326c467 AK |
900 | #undef dirty |
901 | #undef sc | |
902 | #undef unevict | |
903 | #undef mlock | |
2326c467 | 904 | #undef lru |
2326c467 | 905 | #undef head |
2326c467 AK |
906 | #undef slab |
907 | #undef reserved | |
908 | ||
ff604cf6 NH |
909 | /* |
910 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
911 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
912 | */ | |
cc3e2af4 XX |
913 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
914 | enum mf_result result) | |
6a46079c | 915 | { |
97f0b134 XX |
916 | trace_memory_failure_event(pfn, type, result); |
917 | ||
495367c0 | 918 | pr_err("Memory failure: %#lx: recovery action for %s: %s\n", |
64d37a2b | 919 | pfn, action_page_types[type], action_name[result]); |
6a46079c AK |
920 | } |
921 | ||
922 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 923 | unsigned long pfn) |
6a46079c AK |
924 | { |
925 | int result; | |
7456b040 | 926 | int count; |
6a46079c AK |
927 | |
928 | result = ps->action(p, pfn); | |
7456b040 | 929 | |
bd1ce5f9 | 930 | count = page_count(p) - 1; |
cc637b17 | 931 | if (ps->action == me_swapcache_dirty && result == MF_DELAYED) |
138ce286 | 932 | count--; |
78bb9203 | 933 | if (count > 0) { |
495367c0 | 934 | pr_err("Memory failure: %#lx: %s still referenced by %d users\n", |
64d37a2b | 935 | pfn, action_page_types[ps->type], count); |
cc637b17 | 936 | result = MF_FAILED; |
138ce286 | 937 | } |
64d37a2b | 938 | action_result(pfn, ps->type, result); |
6a46079c AK |
939 | |
940 | /* Could do more checks here if page looks ok */ | |
941 | /* | |
942 | * Could adjust zone counters here to correct for the missing page. | |
943 | */ | |
944 | ||
cc637b17 | 945 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
946 | } |
947 | ||
ead07f6a | 948 | /** |
17e395b6 | 949 | * __get_hwpoison_page() - Get refcount for memory error handling: |
ead07f6a NH |
950 | * @page: raw error page (hit by memory error) |
951 | * | |
952 | * Return: return 0 if failed to grab the refcount, otherwise true (some | |
953 | * non-zero value.) | |
954 | */ | |
17e395b6 | 955 | static int __get_hwpoison_page(struct page *page) |
ead07f6a NH |
956 | { |
957 | struct page *head = compound_head(page); | |
958 | ||
4e41a30c | 959 | if (!PageHuge(head) && PageTransHuge(head)) { |
98ed2b00 NH |
960 | /* |
961 | * Non anonymous thp exists only in allocation/free time. We | |
962 | * can't handle such a case correctly, so let's give it up. | |
963 | * This should be better than triggering BUG_ON when kernel | |
964 | * tries to touch the "partially handled" page. | |
965 | */ | |
966 | if (!PageAnon(head)) { | |
495367c0 | 967 | pr_err("Memory failure: %#lx: non anonymous thp\n", |
98ed2b00 NH |
968 | page_to_pfn(page)); |
969 | return 0; | |
970 | } | |
ead07f6a NH |
971 | } |
972 | ||
c2e7e00b KK |
973 | if (get_page_unless_zero(head)) { |
974 | if (head == compound_head(page)) | |
975 | return 1; | |
976 | ||
495367c0 CY |
977 | pr_info("Memory failure: %#lx cannot catch tail\n", |
978 | page_to_pfn(page)); | |
c2e7e00b KK |
979 | put_page(head); |
980 | } | |
981 | ||
982 | return 0; | |
ead07f6a | 983 | } |
ead07f6a | 984 | |
2f714160 OS |
985 | /* |
986 | * Safely get reference count of an arbitrary page. | |
987 | * | |
988 | * Returns 0 for a free page, 1 for an in-use page, | |
989 | * -EIO for a page-type we cannot handle and -EBUSY if we raced with an | |
990 | * allocation. | |
991 | * We only incremented refcount in case the page was already in-use and it | |
992 | * is a known type we can handle. | |
993 | */ | |
994 | static int get_any_page(struct page *p, unsigned long flags) | |
17e395b6 | 995 | { |
2f714160 OS |
996 | int ret = 0, pass = 0; |
997 | bool count_increased = false; | |
17e395b6 | 998 | |
2f714160 OS |
999 | if (flags & MF_COUNT_INCREASED) |
1000 | count_increased = true; | |
1001 | ||
1002 | try_again: | |
1003 | if (!count_increased && !__get_hwpoison_page(p)) { | |
1004 | if (page_count(p)) { | |
1005 | /* We raced with an allocation, retry. */ | |
1006 | if (pass++ < 3) | |
1007 | goto try_again; | |
1008 | ret = -EBUSY; | |
1009 | } else if (!PageHuge(p) && !is_free_buddy_page(p)) { | |
1010 | /* We raced with put_page, retry. */ | |
1011 | if (pass++ < 3) | |
1012 | goto try_again; | |
1013 | ret = -EIO; | |
1014 | } | |
1015 | } else { | |
1016 | if (PageHuge(p) || PageLRU(p) || __PageMovable(p)) { | |
1017 | ret = 1; | |
1018 | } else { | |
1019 | /* | |
1020 | * A page we cannot handle. Check whether we can turn | |
1021 | * it into something we can handle. | |
1022 | */ | |
1023 | if (pass++ < 3) { | |
1024 | put_page(p); | |
1025 | shake_page(p, 1); | |
1026 | count_increased = false; | |
1027 | goto try_again; | |
1028 | } | |
1029 | put_page(p); | |
1030 | ret = -EIO; | |
1031 | } | |
17e395b6 OS |
1032 | } |
1033 | ||
1034 | return ret; | |
1035 | } | |
1036 | ||
2f714160 OS |
1037 | static int get_hwpoison_page(struct page *p, unsigned long flags, |
1038 | enum mf_flags ctxt) | |
1039 | { | |
1040 | int ret; | |
1041 | ||
1042 | zone_pcp_disable(page_zone(p)); | |
1043 | if (ctxt == MF_SOFT_OFFLINE) | |
1044 | ret = get_any_page(p, flags); | |
1045 | else | |
1046 | ret = __get_hwpoison_page(p); | |
1047 | zone_pcp_enable(page_zone(p)); | |
1048 | ||
1049 | return ret; | |
1050 | } | |
1051 | ||
6a46079c AK |
1052 | /* |
1053 | * Do all that is necessary to remove user space mappings. Unmap | |
1054 | * the pages and send SIGBUS to the processes if the data was dirty. | |
1055 | */ | |
666e5a40 | 1056 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
83b57531 | 1057 | int flags, struct page **hpagep) |
6a46079c | 1058 | { |
013339df | 1059 | enum ttu_flags ttu = TTU_IGNORE_MLOCK; |
6a46079c AK |
1060 | struct address_space *mapping; |
1061 | LIST_HEAD(tokill); | |
c0d0381a | 1062 | bool unmap_success = true; |
6751ed65 | 1063 | int kill = 1, forcekill; |
54b9dd14 | 1064 | struct page *hpage = *hpagep; |
286c469a | 1065 | bool mlocked = PageMlocked(hpage); |
6a46079c | 1066 | |
93a9eb39 NH |
1067 | /* |
1068 | * Here we are interested only in user-mapped pages, so skip any | |
1069 | * other types of pages. | |
1070 | */ | |
1071 | if (PageReserved(p) || PageSlab(p)) | |
666e5a40 | 1072 | return true; |
93a9eb39 | 1073 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 1074 | return true; |
6a46079c | 1075 | |
6a46079c AK |
1076 | /* |
1077 | * This check implies we don't kill processes if their pages | |
1078 | * are in the swap cache early. Those are always late kills. | |
1079 | */ | |
7af446a8 | 1080 | if (!page_mapped(hpage)) |
666e5a40 | 1081 | return true; |
1668bfd5 | 1082 | |
52089b14 | 1083 | if (PageKsm(p)) { |
495367c0 | 1084 | pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn); |
666e5a40 | 1085 | return false; |
52089b14 | 1086 | } |
6a46079c AK |
1087 | |
1088 | if (PageSwapCache(p)) { | |
495367c0 CY |
1089 | pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n", |
1090 | pfn); | |
6a46079c AK |
1091 | ttu |= TTU_IGNORE_HWPOISON; |
1092 | } | |
1093 | ||
1094 | /* | |
1095 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1096 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1097 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1098 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1099 | */ |
7af446a8 | 1100 | mapping = page_mapping(hpage); |
6751ed65 | 1101 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
f56753ac | 1102 | mapping_can_writeback(mapping)) { |
7af446a8 NH |
1103 | if (page_mkclean(hpage)) { |
1104 | SetPageDirty(hpage); | |
6a46079c AK |
1105 | } else { |
1106 | kill = 0; | |
1107 | ttu |= TTU_IGNORE_HWPOISON; | |
495367c0 | 1108 | pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1109 | pfn); |
1110 | } | |
1111 | } | |
1112 | ||
1113 | /* | |
1114 | * First collect all the processes that have the page | |
1115 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1116 | * because ttu takes the rmap data structures down. | |
1117 | * | |
1118 | * Error handling: We ignore errors here because | |
1119 | * there's nothing that can be done. | |
1120 | */ | |
1121 | if (kill) | |
415c64c1 | 1122 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1123 | |
c0d0381a MK |
1124 | if (!PageHuge(hpage)) { |
1125 | unmap_success = try_to_unmap(hpage, ttu); | |
1126 | } else { | |
336bf30e MK |
1127 | if (!PageAnon(hpage)) { |
1128 | /* | |
1129 | * For hugetlb pages in shared mappings, try_to_unmap | |
1130 | * could potentially call huge_pmd_unshare. Because of | |
1131 | * this, take semaphore in write mode here and set | |
1132 | * TTU_RMAP_LOCKED to indicate we have taken the lock | |
1133 | * at this higer level. | |
1134 | */ | |
1135 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1136 | if (mapping) { | |
1137 | unmap_success = try_to_unmap(hpage, | |
c0d0381a | 1138 | ttu|TTU_RMAP_LOCKED); |
336bf30e MK |
1139 | i_mmap_unlock_write(mapping); |
1140 | } else { | |
1141 | pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn); | |
1142 | unmap_success = false; | |
1143 | } | |
c0d0381a | 1144 | } else { |
336bf30e | 1145 | unmap_success = try_to_unmap(hpage, ttu); |
c0d0381a MK |
1146 | } |
1147 | } | |
666e5a40 | 1148 | if (!unmap_success) |
495367c0 | 1149 | pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1150 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1151 | |
286c469a NH |
1152 | /* |
1153 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1154 | * shake_page() again to ensure that it's flushed. | |
1155 | */ | |
1156 | if (mlocked) | |
1157 | shake_page(hpage, 0); | |
1158 | ||
6a46079c AK |
1159 | /* |
1160 | * Now that the dirty bit has been propagated to the | |
1161 | * struct page and all unmaps done we can decide if | |
1162 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1163 | * was dirty or the process is not restartable, |
1164 | * otherwise the tokill list is merely | |
6a46079c AK |
1165 | * freed. When there was a problem unmapping earlier |
1166 | * use a more force-full uncatchable kill to prevent | |
1167 | * any accesses to the poisoned memory. | |
1168 | */ | |
415c64c1 | 1169 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
ae1139ec | 1170 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1171 | |
666e5a40 | 1172 | return unmap_success; |
6a46079c AK |
1173 | } |
1174 | ||
0348d2eb NH |
1175 | static int identify_page_state(unsigned long pfn, struct page *p, |
1176 | unsigned long page_flags) | |
761ad8d7 NH |
1177 | { |
1178 | struct page_state *ps; | |
0348d2eb NH |
1179 | |
1180 | /* | |
1181 | * The first check uses the current page flags which may not have any | |
1182 | * relevant information. The second check with the saved page flags is | |
1183 | * carried out only if the first check can't determine the page status. | |
1184 | */ | |
1185 | for (ps = error_states;; ps++) | |
1186 | if ((p->flags & ps->mask) == ps->res) | |
1187 | break; | |
1188 | ||
1189 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1190 | ||
1191 | if (!ps->mask) | |
1192 | for (ps = error_states;; ps++) | |
1193 | if ((page_flags & ps->mask) == ps->res) | |
1194 | break; | |
1195 | return page_action(ps, p, pfn); | |
1196 | } | |
1197 | ||
694bf0b0 OS |
1198 | static int try_to_split_thp_page(struct page *page, const char *msg) |
1199 | { | |
1200 | lock_page(page); | |
1201 | if (!PageAnon(page) || unlikely(split_huge_page(page))) { | |
1202 | unsigned long pfn = page_to_pfn(page); | |
1203 | ||
1204 | unlock_page(page); | |
1205 | if (!PageAnon(page)) | |
1206 | pr_info("%s: %#lx: non anonymous thp\n", msg, pfn); | |
1207 | else | |
1208 | pr_info("%s: %#lx: thp split failed\n", msg, pfn); | |
1209 | put_page(page); | |
1210 | return -EBUSY; | |
1211 | } | |
1212 | unlock_page(page); | |
1213 | ||
1214 | return 0; | |
1215 | } | |
1216 | ||
83b57531 | 1217 | static int memory_failure_hugetlb(unsigned long pfn, int flags) |
0348d2eb | 1218 | { |
761ad8d7 NH |
1219 | struct page *p = pfn_to_page(pfn); |
1220 | struct page *head = compound_head(p); | |
1221 | int res; | |
1222 | unsigned long page_flags; | |
1223 | ||
1224 | if (TestSetPageHWPoison(head)) { | |
1225 | pr_err("Memory failure: %#lx: already hardware poisoned\n", | |
1226 | pfn); | |
1227 | return 0; | |
1228 | } | |
1229 | ||
1230 | num_poisoned_pages_inc(); | |
1231 | ||
2f714160 | 1232 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p, flags, 0)) { |
761ad8d7 NH |
1233 | /* |
1234 | * Check "filter hit" and "race with other subpage." | |
1235 | */ | |
1236 | lock_page(head); | |
1237 | if (PageHWPoison(head)) { | |
1238 | if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1239 | || (p != head && TestSetPageHWPoison(head))) { | |
1240 | num_poisoned_pages_dec(); | |
1241 | unlock_page(head); | |
1242 | return 0; | |
1243 | } | |
1244 | } | |
1245 | unlock_page(head); | |
a8b2c2ce OS |
1246 | res = MF_FAILED; |
1247 | if (!dissolve_free_huge_page(p) && take_page_off_buddy(p)) { | |
1248 | page_ref_inc(p); | |
1249 | res = MF_RECOVERED; | |
1250 | } | |
1251 | action_result(pfn, MF_MSG_FREE_HUGE, res); | |
1252 | return res == MF_RECOVERED ? 0 : -EBUSY; | |
761ad8d7 NH |
1253 | } |
1254 | ||
1255 | lock_page(head); | |
1256 | page_flags = head->flags; | |
1257 | ||
1258 | if (!PageHWPoison(head)) { | |
1259 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); | |
1260 | num_poisoned_pages_dec(); | |
1261 | unlock_page(head); | |
dd6e2402 | 1262 | put_page(head); |
761ad8d7 NH |
1263 | return 0; |
1264 | } | |
1265 | ||
31286a84 NH |
1266 | /* |
1267 | * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so | |
1268 | * simply disable it. In order to make it work properly, we need | |
1269 | * make sure that: | |
1270 | * - conversion of a pud that maps an error hugetlb into hwpoison | |
1271 | * entry properly works, and | |
1272 | * - other mm code walking over page table is aware of pud-aligned | |
1273 | * hwpoison entries. | |
1274 | */ | |
1275 | if (huge_page_size(page_hstate(head)) > PMD_SIZE) { | |
1276 | action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED); | |
1277 | res = -EBUSY; | |
1278 | goto out; | |
1279 | } | |
1280 | ||
83b57531 | 1281 | if (!hwpoison_user_mappings(p, pfn, flags, &head)) { |
761ad8d7 NH |
1282 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1283 | res = -EBUSY; | |
1284 | goto out; | |
1285 | } | |
1286 | ||
0348d2eb | 1287 | res = identify_page_state(pfn, p, page_flags); |
761ad8d7 NH |
1288 | out: |
1289 | unlock_page(head); | |
1290 | return res; | |
1291 | } | |
1292 | ||
6100e34b DW |
1293 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
1294 | struct dev_pagemap *pgmap) | |
1295 | { | |
1296 | struct page *page = pfn_to_page(pfn); | |
1297 | const bool unmap_success = true; | |
1298 | unsigned long size = 0; | |
1299 | struct to_kill *tk; | |
1300 | LIST_HEAD(tokill); | |
1301 | int rc = -EBUSY; | |
1302 | loff_t start; | |
27359fd6 | 1303 | dax_entry_t cookie; |
6100e34b | 1304 | |
1e8aaedb OS |
1305 | if (flags & MF_COUNT_INCREASED) |
1306 | /* | |
1307 | * Drop the extra refcount in case we come from madvise(). | |
1308 | */ | |
1309 | put_page(page); | |
1310 | ||
6100e34b DW |
1311 | /* |
1312 | * Prevent the inode from being freed while we are interrogating | |
1313 | * the address_space, typically this would be handled by | |
1314 | * lock_page(), but dax pages do not use the page lock. This | |
1315 | * also prevents changes to the mapping of this pfn until | |
1316 | * poison signaling is complete. | |
1317 | */ | |
27359fd6 MW |
1318 | cookie = dax_lock_page(page); |
1319 | if (!cookie) | |
6100e34b DW |
1320 | goto out; |
1321 | ||
1322 | if (hwpoison_filter(page)) { | |
1323 | rc = 0; | |
1324 | goto unlock; | |
1325 | } | |
1326 | ||
25b2995a | 1327 | if (pgmap->type == MEMORY_DEVICE_PRIVATE) { |
6100e34b DW |
1328 | /* |
1329 | * TODO: Handle HMM pages which may need coordination | |
1330 | * with device-side memory. | |
1331 | */ | |
1332 | goto unlock; | |
6100e34b DW |
1333 | } |
1334 | ||
1335 | /* | |
1336 | * Use this flag as an indication that the dax page has been | |
1337 | * remapped UC to prevent speculative consumption of poison. | |
1338 | */ | |
1339 | SetPageHWPoison(page); | |
1340 | ||
1341 | /* | |
1342 | * Unlike System-RAM there is no possibility to swap in a | |
1343 | * different physical page at a given virtual address, so all | |
1344 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1345 | * SIGBUS (i.e. MF_MUST_KILL) | |
1346 | */ | |
1347 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1348 | collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED); | |
1349 | ||
1350 | list_for_each_entry(tk, &tokill, nd) | |
1351 | if (tk->size_shift) | |
1352 | size = max(size, 1UL << tk->size_shift); | |
1353 | if (size) { | |
1354 | /* | |
1355 | * Unmap the largest mapping to avoid breaking up | |
1356 | * device-dax mappings which are constant size. The | |
1357 | * actual size of the mapping being torn down is | |
1358 | * communicated in siginfo, see kill_proc() | |
1359 | */ | |
1360 | start = (page->index << PAGE_SHIFT) & ~(size - 1); | |
1361 | unmap_mapping_range(page->mapping, start, start + size, 0); | |
1362 | } | |
1363 | kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags); | |
1364 | rc = 0; | |
1365 | unlock: | |
27359fd6 | 1366 | dax_unlock_page(page, cookie); |
6100e34b DW |
1367 | out: |
1368 | /* drop pgmap ref acquired in caller */ | |
1369 | put_dev_pagemap(pgmap); | |
1370 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
1371 | return rc; | |
1372 | } | |
1373 | ||
cd42f4a3 TL |
1374 | /** |
1375 | * memory_failure - Handle memory failure of a page. | |
1376 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
1377 | * @flags: fine tune action taken |
1378 | * | |
1379 | * This function is called by the low level machine check code | |
1380 | * of an architecture when it detects hardware memory corruption | |
1381 | * of a page. It tries its best to recover, which includes | |
1382 | * dropping pages, killing processes etc. | |
1383 | * | |
1384 | * The function is primarily of use for corruptions that | |
1385 | * happen outside the current execution context (e.g. when | |
1386 | * detected by a background scrubber) | |
1387 | * | |
1388 | * Must run in process context (e.g. a work queue) with interrupts | |
1389 | * enabled and no spinlocks hold. | |
1390 | */ | |
83b57531 | 1391 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 1392 | { |
6a46079c | 1393 | struct page *p; |
7af446a8 | 1394 | struct page *hpage; |
415c64c1 | 1395 | struct page *orig_head; |
6100e34b | 1396 | struct dev_pagemap *pgmap; |
6a46079c | 1397 | int res; |
524fca1e | 1398 | unsigned long page_flags; |
a8b2c2ce | 1399 | bool retry = true; |
6a46079c AK |
1400 | |
1401 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 1402 | panic("Memory failure on page %lx", pfn); |
6a46079c | 1403 | |
96c804a6 DH |
1404 | p = pfn_to_online_page(pfn); |
1405 | if (!p) { | |
1406 | if (pfn_valid(pfn)) { | |
1407 | pgmap = get_dev_pagemap(pfn, NULL); | |
1408 | if (pgmap) | |
1409 | return memory_failure_dev_pagemap(pfn, flags, | |
1410 | pgmap); | |
1411 | } | |
495367c0 CY |
1412 | pr_err("Memory failure: %#lx: memory outside kernel control\n", |
1413 | pfn); | |
a7560fc8 | 1414 | return -ENXIO; |
6a46079c AK |
1415 | } |
1416 | ||
a8b2c2ce | 1417 | try_again: |
761ad8d7 | 1418 | if (PageHuge(p)) |
83b57531 | 1419 | return memory_failure_hugetlb(pfn, flags); |
6a46079c | 1420 | if (TestSetPageHWPoison(p)) { |
495367c0 CY |
1421 | pr_err("Memory failure: %#lx: already hardware poisoned\n", |
1422 | pfn); | |
6a46079c AK |
1423 | return 0; |
1424 | } | |
1425 | ||
761ad8d7 | 1426 | orig_head = hpage = compound_head(p); |
b37ff71c | 1427 | num_poisoned_pages_inc(); |
6a46079c AK |
1428 | |
1429 | /* | |
1430 | * We need/can do nothing about count=0 pages. | |
1431 | * 1) it's a free page, and therefore in safe hand: | |
1432 | * prep_new_page() will be the gate keeper. | |
761ad8d7 | 1433 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
1434 | * Implies some kernel user: cannot stop them from |
1435 | * R/W the page; let's pray that the page has been | |
1436 | * used and will be freed some time later. | |
1437 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 1438 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 1439 | */ |
2f714160 | 1440 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p, flags, 0)) { |
8d22ba1b | 1441 | if (is_free_buddy_page(p)) { |
a8b2c2ce OS |
1442 | if (take_page_off_buddy(p)) { |
1443 | page_ref_inc(p); | |
1444 | res = MF_RECOVERED; | |
1445 | } else { | |
1446 | /* We lost the race, try again */ | |
1447 | if (retry) { | |
1448 | ClearPageHWPoison(p); | |
1449 | num_poisoned_pages_dec(); | |
1450 | retry = false; | |
1451 | goto try_again; | |
1452 | } | |
1453 | res = MF_FAILED; | |
1454 | } | |
1455 | action_result(pfn, MF_MSG_BUDDY, res); | |
1456 | return res == MF_RECOVERED ? 0 : -EBUSY; | |
8d22ba1b | 1457 | } else { |
cc637b17 | 1458 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); |
8d22ba1b WF |
1459 | return -EBUSY; |
1460 | } | |
6a46079c AK |
1461 | } |
1462 | ||
761ad8d7 | 1463 | if (PageTransHuge(hpage)) { |
5d1fd5dc NH |
1464 | if (try_to_split_thp_page(p, "Memory Failure") < 0) { |
1465 | action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED); | |
415c64c1 | 1466 | return -EBUSY; |
5d1fd5dc | 1467 | } |
415c64c1 | 1468 | VM_BUG_ON_PAGE(!page_count(p), p); |
415c64c1 NH |
1469 | } |
1470 | ||
e43c3afb WF |
1471 | /* |
1472 | * We ignore non-LRU pages for good reasons. | |
1473 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 1474 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
1475 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
1476 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1477 | * walked by the page reclaim code, however that's not a big loss. | |
1478 | */ | |
8bcb74de | 1479 | shake_page(p, 0); |
e43c3afb | 1480 | |
761ad8d7 | 1481 | lock_page(p); |
847ce401 | 1482 | |
f37d4298 AK |
1483 | /* |
1484 | * The page could have changed compound pages during the locking. | |
1485 | * If this happens just bail out. | |
1486 | */ | |
415c64c1 | 1487 | if (PageCompound(p) && compound_head(p) != orig_head) { |
cc637b17 | 1488 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 AK |
1489 | res = -EBUSY; |
1490 | goto out; | |
1491 | } | |
1492 | ||
524fca1e NH |
1493 | /* |
1494 | * We use page flags to determine what action should be taken, but | |
1495 | * the flags can be modified by the error containment action. One | |
1496 | * example is an mlocked page, where PG_mlocked is cleared by | |
1497 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1498 | * correctly, we save a copy of the page flags at this time. | |
1499 | */ | |
7d9d46ac | 1500 | page_flags = p->flags; |
524fca1e | 1501 | |
847ce401 WF |
1502 | /* |
1503 | * unpoison always clear PG_hwpoison inside page lock | |
1504 | */ | |
1505 | if (!PageHWPoison(p)) { | |
495367c0 | 1506 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); |
b37ff71c | 1507 | num_poisoned_pages_dec(); |
761ad8d7 | 1508 | unlock_page(p); |
dd6e2402 | 1509 | put_page(p); |
a09233f3 | 1510 | return 0; |
847ce401 | 1511 | } |
7c116f2b WF |
1512 | if (hwpoison_filter(p)) { |
1513 | if (TestClearPageHWPoison(p)) | |
b37ff71c | 1514 | num_poisoned_pages_dec(); |
761ad8d7 | 1515 | unlock_page(p); |
dd6e2402 | 1516 | put_page(p); |
7c116f2b WF |
1517 | return 0; |
1518 | } | |
847ce401 | 1519 | |
761ad8d7 | 1520 | if (!PageTransTail(p) && !PageLRU(p)) |
0bc1f8b0 CY |
1521 | goto identify_page_state; |
1522 | ||
6edd6cc6 NH |
1523 | /* |
1524 | * It's very difficult to mess with pages currently under IO | |
1525 | * and in many cases impossible, so we just avoid it here. | |
1526 | */ | |
6a46079c AK |
1527 | wait_on_page_writeback(p); |
1528 | ||
1529 | /* | |
1530 | * Now take care of user space mappings. | |
e64a782f | 1531 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
6a46079c | 1532 | */ |
1b473bec | 1533 | if (!hwpoison_user_mappings(p, pfn, flags, &p)) { |
cc637b17 | 1534 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 WF |
1535 | res = -EBUSY; |
1536 | goto out; | |
1537 | } | |
6a46079c AK |
1538 | |
1539 | /* | |
1540 | * Torn down by someone else? | |
1541 | */ | |
dc2a1cbf | 1542 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 1543 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 1544 | res = -EBUSY; |
6a46079c AK |
1545 | goto out; |
1546 | } | |
1547 | ||
0bc1f8b0 | 1548 | identify_page_state: |
0348d2eb | 1549 | res = identify_page_state(pfn, p, page_flags); |
6a46079c | 1550 | out: |
761ad8d7 | 1551 | unlock_page(p); |
6a46079c AK |
1552 | return res; |
1553 | } | |
cd42f4a3 | 1554 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1555 | |
ea8f5fb8 HY |
1556 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1557 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1558 | ||
1559 | struct memory_failure_entry { | |
1560 | unsigned long pfn; | |
ea8f5fb8 HY |
1561 | int flags; |
1562 | }; | |
1563 | ||
1564 | struct memory_failure_cpu { | |
1565 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1566 | MEMORY_FAILURE_FIFO_SIZE); | |
1567 | spinlock_t lock; | |
1568 | struct work_struct work; | |
1569 | }; | |
1570 | ||
1571 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1572 | ||
1573 | /** | |
1574 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1575 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
1576 | * @flags: Flags for memory failure handling |
1577 | * | |
1578 | * This function is called by the low level hardware error handler | |
1579 | * when it detects hardware memory corruption of a page. It schedules | |
1580 | * the recovering of error page, including dropping pages, killing | |
1581 | * processes etc. | |
1582 | * | |
1583 | * The function is primarily of use for corruptions that | |
1584 | * happen outside the current execution context (e.g. when | |
1585 | * detected by a background scrubber) | |
1586 | * | |
1587 | * Can run in IRQ context. | |
1588 | */ | |
83b57531 | 1589 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
1590 | { |
1591 | struct memory_failure_cpu *mf_cpu; | |
1592 | unsigned long proc_flags; | |
1593 | struct memory_failure_entry entry = { | |
1594 | .pfn = pfn, | |
ea8f5fb8 HY |
1595 | .flags = flags, |
1596 | }; | |
1597 | ||
1598 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1599 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 1600 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
1601 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
1602 | else | |
8e33a52f | 1603 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
1604 | pfn); |
1605 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1606 | put_cpu_var(memory_failure_cpu); | |
1607 | } | |
1608 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1609 | ||
1610 | static void memory_failure_work_func(struct work_struct *work) | |
1611 | { | |
1612 | struct memory_failure_cpu *mf_cpu; | |
1613 | struct memory_failure_entry entry = { 0, }; | |
1614 | unsigned long proc_flags; | |
1615 | int gotten; | |
1616 | ||
06202231 | 1617 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
ea8f5fb8 HY |
1618 | for (;;) { |
1619 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1620 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1621 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1622 | if (!gotten) | |
1623 | break; | |
cf870c70 | 1624 | if (entry.flags & MF_SOFT_OFFLINE) |
feec24a6 | 1625 | soft_offline_page(entry.pfn, entry.flags); |
cf870c70 | 1626 | else |
83b57531 | 1627 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
1628 | } |
1629 | } | |
1630 | ||
06202231 JM |
1631 | /* |
1632 | * Process memory_failure work queued on the specified CPU. | |
1633 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. | |
1634 | */ | |
1635 | void memory_failure_queue_kick(int cpu) | |
1636 | { | |
1637 | struct memory_failure_cpu *mf_cpu; | |
1638 | ||
1639 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1640 | cancel_work_sync(&mf_cpu->work); | |
1641 | memory_failure_work_func(&mf_cpu->work); | |
1642 | } | |
1643 | ||
ea8f5fb8 HY |
1644 | static int __init memory_failure_init(void) |
1645 | { | |
1646 | struct memory_failure_cpu *mf_cpu; | |
1647 | int cpu; | |
1648 | ||
1649 | for_each_possible_cpu(cpu) { | |
1650 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1651 | spin_lock_init(&mf_cpu->lock); | |
1652 | INIT_KFIFO(mf_cpu->fifo); | |
1653 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1654 | } | |
1655 | ||
1656 | return 0; | |
1657 | } | |
1658 | core_initcall(memory_failure_init); | |
1659 | ||
a5f65109 NH |
1660 | #define unpoison_pr_info(fmt, pfn, rs) \ |
1661 | ({ \ | |
1662 | if (__ratelimit(rs)) \ | |
1663 | pr_info(fmt, pfn); \ | |
1664 | }) | |
1665 | ||
847ce401 WF |
1666 | /** |
1667 | * unpoison_memory - Unpoison a previously poisoned page | |
1668 | * @pfn: Page number of the to be unpoisoned page | |
1669 | * | |
1670 | * Software-unpoison a page that has been poisoned by | |
1671 | * memory_failure() earlier. | |
1672 | * | |
1673 | * This is only done on the software-level, so it only works | |
1674 | * for linux injected failures, not real hardware failures | |
1675 | * | |
1676 | * Returns 0 for success, otherwise -errno. | |
1677 | */ | |
1678 | int unpoison_memory(unsigned long pfn) | |
1679 | { | |
1680 | struct page *page; | |
1681 | struct page *p; | |
1682 | int freeit = 0; | |
2f714160 | 1683 | unsigned long flags = 0; |
a5f65109 NH |
1684 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
1685 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
1686 | |
1687 | if (!pfn_valid(pfn)) | |
1688 | return -ENXIO; | |
1689 | ||
1690 | p = pfn_to_page(pfn); | |
1691 | page = compound_head(p); | |
1692 | ||
1693 | if (!PageHWPoison(p)) { | |
495367c0 | 1694 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 1695 | pfn, &unpoison_rs); |
847ce401 WF |
1696 | return 0; |
1697 | } | |
1698 | ||
230ac719 | 1699 | if (page_count(page) > 1) { |
495367c0 | 1700 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 1701 | pfn, &unpoison_rs); |
230ac719 NH |
1702 | return 0; |
1703 | } | |
1704 | ||
1705 | if (page_mapped(page)) { | |
495367c0 | 1706 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 1707 | pfn, &unpoison_rs); |
230ac719 NH |
1708 | return 0; |
1709 | } | |
1710 | ||
1711 | if (page_mapping(page)) { | |
495367c0 | 1712 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 1713 | pfn, &unpoison_rs); |
230ac719 NH |
1714 | return 0; |
1715 | } | |
1716 | ||
0cea3fdc WL |
1717 | /* |
1718 | * unpoison_memory() can encounter thp only when the thp is being | |
1719 | * worked by memory_failure() and the page lock is not held yet. | |
1720 | * In such case, we yield to memory_failure() and make unpoison fail. | |
1721 | */ | |
e76d30e2 | 1722 | if (!PageHuge(page) && PageTransHuge(page)) { |
495367c0 | 1723 | unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n", |
a5f65109 | 1724 | pfn, &unpoison_rs); |
ead07f6a | 1725 | return 0; |
0cea3fdc WL |
1726 | } |
1727 | ||
2f714160 | 1728 | if (!get_hwpoison_page(p, flags, 0)) { |
847ce401 | 1729 | if (TestClearPageHWPoison(p)) |
8e30456b | 1730 | num_poisoned_pages_dec(); |
495367c0 | 1731 | unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n", |
a5f65109 | 1732 | pfn, &unpoison_rs); |
847ce401 WF |
1733 | return 0; |
1734 | } | |
1735 | ||
7eaceacc | 1736 | lock_page(page); |
847ce401 WF |
1737 | /* |
1738 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1739 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1740 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1741 | * the free buddy page pool. | |
1742 | */ | |
c9fbdd5f | 1743 | if (TestClearPageHWPoison(page)) { |
495367c0 | 1744 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", |
a5f65109 | 1745 | pfn, &unpoison_rs); |
b37ff71c | 1746 | num_poisoned_pages_dec(); |
847ce401 WF |
1747 | freeit = 1; |
1748 | } | |
1749 | unlock_page(page); | |
1750 | ||
dd6e2402 | 1751 | put_page(page); |
3ba5eebc | 1752 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) |
dd6e2402 | 1753 | put_page(page); |
847ce401 WF |
1754 | |
1755 | return 0; | |
1756 | } | |
1757 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 1758 | |
6b9a217e | 1759 | static bool isolate_page(struct page *page, struct list_head *pagelist) |
d950b958 | 1760 | { |
6b9a217e OS |
1761 | bool isolated = false; |
1762 | bool lru = PageLRU(page); | |
d950b958 | 1763 | |
6b9a217e OS |
1764 | if (PageHuge(page)) { |
1765 | isolated = isolate_huge_page(page, pagelist); | |
1766 | } else { | |
1767 | if (lru) | |
1768 | isolated = !isolate_lru_page(page); | |
1769 | else | |
1770 | isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE); | |
1771 | ||
1772 | if (isolated) | |
1773 | list_add(&page->lru, pagelist); | |
0ebff32c | 1774 | } |
d950b958 | 1775 | |
6b9a217e OS |
1776 | if (isolated && lru) |
1777 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
1778 | page_is_file_lru(page)); | |
1779 | ||
03613808 | 1780 | /* |
6b9a217e OS |
1781 | * If we succeed to isolate the page, we grabbed another refcount on |
1782 | * the page, so we can safely drop the one we got from get_any_pages(). | |
1783 | * If we failed to isolate the page, it means that we cannot go further | |
1784 | * and we will return an error, so drop the reference we got from | |
1785 | * get_any_pages() as well. | |
03613808 | 1786 | */ |
6b9a217e OS |
1787 | put_page(page); |
1788 | return isolated; | |
d950b958 NH |
1789 | } |
1790 | ||
6b9a217e OS |
1791 | /* |
1792 | * __soft_offline_page handles hugetlb-pages and non-hugetlb pages. | |
1793 | * If the page is a non-dirty unmapped page-cache page, it simply invalidates. | |
1794 | * If the page is mapped, it migrates the contents over. | |
1795 | */ | |
1796 | static int __soft_offline_page(struct page *page) | |
af8fae7c | 1797 | { |
6b9a217e | 1798 | int ret = 0; |
af8fae7c | 1799 | unsigned long pfn = page_to_pfn(page); |
6b9a217e OS |
1800 | struct page *hpage = compound_head(page); |
1801 | char const *msg_page[] = {"page", "hugepage"}; | |
1802 | bool huge = PageHuge(page); | |
1803 | LIST_HEAD(pagelist); | |
54608759 JK |
1804 | struct migration_target_control mtc = { |
1805 | .nid = NUMA_NO_NODE, | |
1806 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
1807 | }; | |
facb6011 | 1808 | |
facb6011 | 1809 | /* |
af8fae7c NH |
1810 | * Check PageHWPoison again inside page lock because PageHWPoison |
1811 | * is set by memory_failure() outside page lock. Note that | |
1812 | * memory_failure() also double-checks PageHWPoison inside page lock, | |
1813 | * so there's no race between soft_offline_page() and memory_failure(). | |
facb6011 | 1814 | */ |
0ebff32c | 1815 | lock_page(page); |
6b9a217e OS |
1816 | if (!PageHuge(page)) |
1817 | wait_on_page_writeback(page); | |
af8fae7c NH |
1818 | if (PageHWPoison(page)) { |
1819 | unlock_page(page); | |
dd6e2402 | 1820 | put_page(page); |
af8fae7c | 1821 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
5a2ffca3 | 1822 | return 0; |
af8fae7c | 1823 | } |
6b9a217e OS |
1824 | |
1825 | if (!PageHuge(page)) | |
1826 | /* | |
1827 | * Try to invalidate first. This should work for | |
1828 | * non dirty unmapped page cache pages. | |
1829 | */ | |
1830 | ret = invalidate_inode_page(page); | |
facb6011 | 1831 | unlock_page(page); |
6b9a217e | 1832 | |
facb6011 | 1833 | /* |
facb6011 AK |
1834 | * RED-PEN would be better to keep it isolated here, but we |
1835 | * would need to fix isolation locking first. | |
1836 | */ | |
6b9a217e | 1837 | if (ret) { |
fb46e735 | 1838 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
6b9a217e | 1839 | page_handle_poison(page, false, true); |
af8fae7c | 1840 | return 0; |
facb6011 AK |
1841 | } |
1842 | ||
6b9a217e | 1843 | if (isolate_page(hpage, &pagelist)) { |
54608759 JK |
1844 | ret = migrate_pages(&pagelist, alloc_migration_target, NULL, |
1845 | (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE); | |
79f5f8fa | 1846 | if (!ret) { |
6b9a217e OS |
1847 | bool release = !huge; |
1848 | ||
1849 | if (!page_handle_poison(page, huge, release)) | |
1850 | ret = -EBUSY; | |
79f5f8fa | 1851 | } else { |
85fbe5d1 YX |
1852 | if (!list_empty(&pagelist)) |
1853 | putback_movable_pages(&pagelist); | |
59c82b70 | 1854 | |
6b9a217e OS |
1855 | pr_info("soft offline: %#lx: %s migration failed %d, type %lx (%pGp)\n", |
1856 | pfn, msg_page[huge], ret, page->flags, &page->flags); | |
facb6011 | 1857 | if (ret > 0) |
3f4b815a | 1858 | ret = -EBUSY; |
facb6011 AK |
1859 | } |
1860 | } else { | |
3f4b815a OS |
1861 | pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %lx (%pGp)\n", |
1862 | pfn, msg_page[huge], page_count(page), page->flags, &page->flags); | |
6b9a217e | 1863 | ret = -EBUSY; |
facb6011 | 1864 | } |
facb6011 AK |
1865 | return ret; |
1866 | } | |
86e05773 | 1867 | |
6b9a217e | 1868 | static int soft_offline_in_use_page(struct page *page) |
acc14dc4 | 1869 | { |
acc14dc4 NH |
1870 | struct page *hpage = compound_head(page); |
1871 | ||
694bf0b0 OS |
1872 | if (!PageHuge(page) && PageTransHuge(hpage)) |
1873 | if (try_to_split_thp_page(page, "soft offline") < 0) | |
acc14dc4 | 1874 | return -EBUSY; |
6b9a217e | 1875 | return __soft_offline_page(page); |
acc14dc4 NH |
1876 | } |
1877 | ||
d4ae9916 | 1878 | static int soft_offline_free_page(struct page *page) |
acc14dc4 | 1879 | { |
6b9a217e | 1880 | int rc = 0; |
acc14dc4 | 1881 | |
6b9a217e OS |
1882 | if (!page_handle_poison(page, true, false)) |
1883 | rc = -EBUSY; | |
06be6ff3 | 1884 | |
d4ae9916 | 1885 | return rc; |
acc14dc4 NH |
1886 | } |
1887 | ||
86e05773 WL |
1888 | /** |
1889 | * soft_offline_page - Soft offline a page. | |
feec24a6 | 1890 | * @pfn: pfn to soft-offline |
86e05773 WL |
1891 | * @flags: flags. Same as memory_failure(). |
1892 | * | |
1893 | * Returns 0 on success, otherwise negated errno. | |
1894 | * | |
1895 | * Soft offline a page, by migration or invalidation, | |
1896 | * without killing anything. This is for the case when | |
1897 | * a page is not corrupted yet (so it's still valid to access), | |
1898 | * but has had a number of corrected errors and is better taken | |
1899 | * out. | |
1900 | * | |
1901 | * The actual policy on when to do that is maintained by | |
1902 | * user space. | |
1903 | * | |
1904 | * This should never impact any application or cause data loss, | |
1905 | * however it might take some time. | |
1906 | * | |
1907 | * This is not a 100% solution for all memory, but tries to be | |
1908 | * ``good enough'' for the majority of memory. | |
1909 | */ | |
feec24a6 | 1910 | int soft_offline_page(unsigned long pfn, int flags) |
86e05773 WL |
1911 | { |
1912 | int ret; | |
feec24a6 | 1913 | struct page *page; |
b94e0282 | 1914 | bool try_again = true; |
86e05773 | 1915 | |
feec24a6 NH |
1916 | if (!pfn_valid(pfn)) |
1917 | return -ENXIO; | |
1918 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ | |
1919 | page = pfn_to_online_page(pfn); | |
1920 | if (!page) | |
86a66810 | 1921 | return -EIO; |
86a66810 | 1922 | |
86e05773 | 1923 | if (PageHWPoison(page)) { |
8295d535 | 1924 | pr_info("%s: %#lx page already poisoned\n", __func__, pfn); |
1e0e635b | 1925 | if (flags & MF_COUNT_INCREASED) |
dd6e2402 | 1926 | put_page(page); |
5a2ffca3 | 1927 | return 0; |
86e05773 | 1928 | } |
86e05773 | 1929 | |
b94e0282 | 1930 | retry: |
bfc8c901 | 1931 | get_online_mems(); |
2f714160 | 1932 | ret = get_hwpoison_page(page, flags, MF_SOFT_OFFLINE); |
bfc8c901 | 1933 | put_online_mems(); |
4e41a30c | 1934 | |
8295d535 | 1935 | if (ret > 0) { |
6b9a217e | 1936 | ret = soft_offline_in_use_page(page); |
8295d535 | 1937 | } else if (ret == 0) { |
b94e0282 OS |
1938 | if (soft_offline_free_page(page) && try_again) { |
1939 | try_again = false; | |
1940 | goto retry; | |
1941 | } | |
8295d535 OS |
1942 | } else if (ret == -EIO) { |
1943 | pr_info("%s: %#lx: unknown page type: %lx (%pGP)\n", | |
1944 | __func__, pfn, page->flags, &page->flags); | |
1945 | } | |
4e41a30c | 1946 | |
86e05773 WL |
1947 | return ret; |
1948 | } |