Commit | Line | Data |
---|---|---|
6a46079c AK |
1 | /* |
2 | * Copyright (C) 2008, 2009 Intel Corporation | |
3 | * Authors: Andi Kleen, Fengguang Wu | |
4 | * | |
5 | * This software may be redistributed and/or modified under the terms of | |
6 | * the GNU General Public License ("GPL") version 2 only as published by the | |
7 | * Free Software Foundation. | |
8 | * | |
9 | * High level machine check handler. Handles pages reported by the | |
1c80b990 | 10 | * hardware as being corrupted usually due to a multi-bit ECC memory or cache |
6a46079c | 11 | * failure. |
1c80b990 AK |
12 | * |
13 | * In addition there is a "soft offline" entry point that allows stop using | |
14 | * not-yet-corrupted-by-suspicious pages without killing anything. | |
6a46079c AK |
15 | * |
16 | * Handles page cache pages in various states. The tricky part | |
1c80b990 AK |
17 | * here is that we can access any page asynchronously in respect to |
18 | * other VM users, because memory failures could happen anytime and | |
19 | * anywhere. This could violate some of their assumptions. This is why | |
20 | * this code has to be extremely careful. Generally it tries to use | |
21 | * normal locking rules, as in get the standard locks, even if that means | |
22 | * the error handling takes potentially a long time. | |
e0de78df AK |
23 | * |
24 | * It can be very tempting to add handling for obscure cases here. | |
25 | * In general any code for handling new cases should only be added iff: | |
26 | * - You know how to test it. | |
27 | * - You have a test that can be added to mce-test | |
28 | * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/ | |
29 | * - The case actually shows up as a frequent (top 10) page state in | |
30 | * tools/vm/page-types when running a real workload. | |
1c80b990 AK |
31 | * |
32 | * There are several operations here with exponential complexity because | |
33 | * of unsuitable VM data structures. For example the operation to map back | |
34 | * from RMAP chains to processes has to walk the complete process list and | |
35 | * has non linear complexity with the number. But since memory corruptions | |
36 | * are rare we hope to get away with this. This avoids impacting the core | |
37 | * VM. | |
6a46079c | 38 | */ |
6a46079c AK |
39 | #include <linux/kernel.h> |
40 | #include <linux/mm.h> | |
41 | #include <linux/page-flags.h> | |
478c5ffc | 42 | #include <linux/kernel-page-flags.h> |
6a46079c | 43 | #include <linux/sched.h> |
01e00f88 | 44 | #include <linux/ksm.h> |
6a46079c | 45 | #include <linux/rmap.h> |
b9e15baf | 46 | #include <linux/export.h> |
6a46079c AK |
47 | #include <linux/pagemap.h> |
48 | #include <linux/swap.h> | |
49 | #include <linux/backing-dev.h> | |
facb6011 AK |
50 | #include <linux/migrate.h> |
51 | #include <linux/page-isolation.h> | |
52 | #include <linux/suspend.h> | |
5a0e3ad6 | 53 | #include <linux/slab.h> |
bf998156 | 54 | #include <linux/swapops.h> |
7af446a8 | 55 | #include <linux/hugetlb.h> |
20d6c96b | 56 | #include <linux/memory_hotplug.h> |
5db8a73a | 57 | #include <linux/mm_inline.h> |
ea8f5fb8 | 58 | #include <linux/kfifo.h> |
a5f65109 | 59 | #include <linux/ratelimit.h> |
6a46079c | 60 | #include "internal.h" |
97f0b134 | 61 | #include "ras/ras_event.h" |
6a46079c AK |
62 | |
63 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
64 | ||
65 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
66 | ||
293c07e3 | 67 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 68 | |
27df5068 AK |
69 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
70 | ||
1bfe5feb | 71 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
72 | u32 hwpoison_filter_dev_major = ~0U; |
73 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
74 | u64 hwpoison_filter_flags_mask; |
75 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 76 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
77 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
78 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
79 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
80 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
81 | |
82 | static int hwpoison_filter_dev(struct page *p) | |
83 | { | |
84 | struct address_space *mapping; | |
85 | dev_t dev; | |
86 | ||
87 | if (hwpoison_filter_dev_major == ~0U && | |
88 | hwpoison_filter_dev_minor == ~0U) | |
89 | return 0; | |
90 | ||
91 | /* | |
1c80b990 | 92 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
93 | */ |
94 | if (PageSlab(p)) | |
95 | return -EINVAL; | |
96 | ||
97 | mapping = page_mapping(p); | |
98 | if (mapping == NULL || mapping->host == NULL) | |
99 | return -EINVAL; | |
100 | ||
101 | dev = mapping->host->i_sb->s_dev; | |
102 | if (hwpoison_filter_dev_major != ~0U && | |
103 | hwpoison_filter_dev_major != MAJOR(dev)) | |
104 | return -EINVAL; | |
105 | if (hwpoison_filter_dev_minor != ~0U && | |
106 | hwpoison_filter_dev_minor != MINOR(dev)) | |
107 | return -EINVAL; | |
108 | ||
109 | return 0; | |
110 | } | |
111 | ||
478c5ffc WF |
112 | static int hwpoison_filter_flags(struct page *p) |
113 | { | |
114 | if (!hwpoison_filter_flags_mask) | |
115 | return 0; | |
116 | ||
117 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
118 | hwpoison_filter_flags_value) | |
119 | return 0; | |
120 | else | |
121 | return -EINVAL; | |
122 | } | |
123 | ||
4fd466eb AK |
124 | /* |
125 | * This allows stress tests to limit test scope to a collection of tasks | |
126 | * by putting them under some memcg. This prevents killing unrelated/important | |
127 | * processes such as /sbin/init. Note that the target task may share clean | |
128 | * pages with init (eg. libc text), which is harmless. If the target task | |
129 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
130 | * is also included in the memcg. At last, due to race conditions this filter | |
131 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
132 | * a freed page. | |
133 | */ | |
94a59fb3 | 134 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
135 | u64 hwpoison_filter_memcg; |
136 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
137 | static int hwpoison_filter_task(struct page *p) | |
138 | { | |
4fd466eb AK |
139 | if (!hwpoison_filter_memcg) |
140 | return 0; | |
141 | ||
94a59fb3 | 142 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
143 | return -EINVAL; |
144 | ||
145 | return 0; | |
146 | } | |
147 | #else | |
148 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
149 | #endif | |
150 | ||
7c116f2b WF |
151 | int hwpoison_filter(struct page *p) |
152 | { | |
1bfe5feb HL |
153 | if (!hwpoison_filter_enable) |
154 | return 0; | |
155 | ||
7c116f2b WF |
156 | if (hwpoison_filter_dev(p)) |
157 | return -EINVAL; | |
158 | ||
478c5ffc WF |
159 | if (hwpoison_filter_flags(p)) |
160 | return -EINVAL; | |
161 | ||
4fd466eb AK |
162 | if (hwpoison_filter_task(p)) |
163 | return -EINVAL; | |
164 | ||
7c116f2b WF |
165 | return 0; |
166 | } | |
27df5068 AK |
167 | #else |
168 | int hwpoison_filter(struct page *p) | |
169 | { | |
170 | return 0; | |
171 | } | |
172 | #endif | |
173 | ||
7c116f2b WF |
174 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
175 | ||
6a46079c | 176 | /* |
7329bbeb TL |
177 | * Send all the processes who have the page mapped a signal. |
178 | * ``action optional'' if they are not immediately affected by the error | |
179 | * ``action required'' if error happened in current execution context | |
6a46079c | 180 | */ |
7329bbeb TL |
181 | static int kill_proc(struct task_struct *t, unsigned long addr, int trapno, |
182 | unsigned long pfn, struct page *page, int flags) | |
6a46079c AK |
183 | { |
184 | struct siginfo si; | |
185 | int ret; | |
186 | ||
1170532b JP |
187 | pr_err("MCE %#lx: Killing %s:%d due to hardware memory corruption\n", |
188 | pfn, t->comm, t->pid); | |
6a46079c AK |
189 | si.si_signo = SIGBUS; |
190 | si.si_errno = 0; | |
6a46079c AK |
191 | si.si_addr = (void *)addr; |
192 | #ifdef __ARCH_SI_TRAPNO | |
193 | si.si_trapno = trapno; | |
194 | #endif | |
f9121153 | 195 | si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT; |
7329bbeb | 196 | |
a70ffcac | 197 | if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) { |
7329bbeb | 198 | si.si_code = BUS_MCEERR_AR; |
a70ffcac | 199 | ret = force_sig_info(SIGBUS, &si, current); |
7329bbeb TL |
200 | } else { |
201 | /* | |
202 | * Don't use force here, it's convenient if the signal | |
203 | * can be temporarily blocked. | |
204 | * This could cause a loop when the user sets SIGBUS | |
205 | * to SIG_IGN, but hopefully no one will do that? | |
206 | */ | |
207 | si.si_code = BUS_MCEERR_AO; | |
208 | ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */ | |
209 | } | |
6a46079c | 210 | if (ret < 0) |
1170532b JP |
211 | pr_info("MCE: Error sending signal to %s:%d: %d\n", |
212 | t->comm, t->pid, ret); | |
6a46079c AK |
213 | return ret; |
214 | } | |
215 | ||
588f9ce6 AK |
216 | /* |
217 | * When a unknown page type is encountered drain as many buffers as possible | |
218 | * in the hope to turn the page into a LRU or free page, which we can handle. | |
219 | */ | |
facb6011 | 220 | void shake_page(struct page *p, int access) |
588f9ce6 AK |
221 | { |
222 | if (!PageSlab(p)) { | |
223 | lru_add_drain_all(); | |
224 | if (PageLRU(p)) | |
225 | return; | |
c0554329 | 226 | drain_all_pages(page_zone(p)); |
588f9ce6 AK |
227 | if (PageLRU(p) || is_free_buddy_page(p)) |
228 | return; | |
229 | } | |
facb6011 | 230 | |
588f9ce6 | 231 | /* |
6b4f7799 JW |
232 | * Only call shrink_node_slabs here (which would also shrink |
233 | * other caches) if access is not potentially fatal. | |
588f9ce6 | 234 | */ |
cb731d6c VD |
235 | if (access) |
236 | drop_slab_node(page_to_nid(p)); | |
588f9ce6 AK |
237 | } |
238 | EXPORT_SYMBOL_GPL(shake_page); | |
239 | ||
6a46079c AK |
240 | /* |
241 | * Kill all processes that have a poisoned page mapped and then isolate | |
242 | * the page. | |
243 | * | |
244 | * General strategy: | |
245 | * Find all processes having the page mapped and kill them. | |
246 | * But we keep a page reference around so that the page is not | |
247 | * actually freed yet. | |
248 | * Then stash the page away | |
249 | * | |
250 | * There's no convenient way to get back to mapped processes | |
251 | * from the VMAs. So do a brute-force search over all | |
252 | * running processes. | |
253 | * | |
254 | * Remember that machine checks are not common (or rather | |
255 | * if they are common you have other problems), so this shouldn't | |
256 | * be a performance issue. | |
257 | * | |
258 | * Also there are some races possible while we get from the | |
259 | * error detection to actually handle it. | |
260 | */ | |
261 | ||
262 | struct to_kill { | |
263 | struct list_head nd; | |
264 | struct task_struct *tsk; | |
265 | unsigned long addr; | |
9033ae16 | 266 | char addr_valid; |
6a46079c AK |
267 | }; |
268 | ||
269 | /* | |
270 | * Failure handling: if we can't find or can't kill a process there's | |
271 | * not much we can do. We just print a message and ignore otherwise. | |
272 | */ | |
273 | ||
274 | /* | |
275 | * Schedule a process for later kill. | |
276 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
277 | * TBD would GFP_NOIO be enough? | |
278 | */ | |
279 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
280 | struct vm_area_struct *vma, | |
281 | struct list_head *to_kill, | |
282 | struct to_kill **tkc) | |
283 | { | |
284 | struct to_kill *tk; | |
285 | ||
286 | if (*tkc) { | |
287 | tk = *tkc; | |
288 | *tkc = NULL; | |
289 | } else { | |
290 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); | |
291 | if (!tk) { | |
1170532b | 292 | pr_err("MCE: Out of memory while machine check handling\n"); |
6a46079c AK |
293 | return; |
294 | } | |
295 | } | |
296 | tk->addr = page_address_in_vma(p, vma); | |
297 | tk->addr_valid = 1; | |
298 | ||
299 | /* | |
300 | * In theory we don't have to kill when the page was | |
301 | * munmaped. But it could be also a mremap. Since that's | |
302 | * likely very rare kill anyways just out of paranoia, but use | |
303 | * a SIGKILL because the error is not contained anymore. | |
304 | */ | |
305 | if (tk->addr == -EFAULT) { | |
fb46e735 | 306 | pr_info("MCE: Unable to find user space address %lx in %s\n", |
6a46079c AK |
307 | page_to_pfn(p), tsk->comm); |
308 | tk->addr_valid = 0; | |
309 | } | |
310 | get_task_struct(tsk); | |
311 | tk->tsk = tsk; | |
312 | list_add_tail(&tk->nd, to_kill); | |
313 | } | |
314 | ||
315 | /* | |
316 | * Kill the processes that have been collected earlier. | |
317 | * | |
318 | * Only do anything when DOIT is set, otherwise just free the list | |
319 | * (this is used for clean pages which do not need killing) | |
320 | * Also when FAIL is set do a force kill because something went | |
321 | * wrong earlier. | |
322 | */ | |
6751ed65 | 323 | static void kill_procs(struct list_head *to_kill, int forcekill, int trapno, |
7329bbeb TL |
324 | int fail, struct page *page, unsigned long pfn, |
325 | int flags) | |
6a46079c AK |
326 | { |
327 | struct to_kill *tk, *next; | |
328 | ||
329 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 330 | if (forcekill) { |
6a46079c | 331 | /* |
af901ca1 | 332 | * In case something went wrong with munmapping |
6a46079c AK |
333 | * make sure the process doesn't catch the |
334 | * signal and then access the memory. Just kill it. | |
6a46079c AK |
335 | */ |
336 | if (fail || tk->addr_valid == 0) { | |
1170532b JP |
337 | pr_err("MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
338 | pfn, tk->tsk->comm, tk->tsk->pid); | |
6a46079c AK |
339 | force_sig(SIGKILL, tk->tsk); |
340 | } | |
341 | ||
342 | /* | |
343 | * In theory the process could have mapped | |
344 | * something else on the address in-between. We could | |
345 | * check for that, but we need to tell the | |
346 | * process anyways. | |
347 | */ | |
7329bbeb TL |
348 | else if (kill_proc(tk->tsk, tk->addr, trapno, |
349 | pfn, page, flags) < 0) | |
1170532b JP |
350 | pr_err("MCE %#lx: Cannot send advisory machine check signal to %s:%d\n", |
351 | pfn, tk->tsk->comm, tk->tsk->pid); | |
6a46079c AK |
352 | } |
353 | put_task_struct(tk->tsk); | |
354 | kfree(tk); | |
355 | } | |
356 | } | |
357 | ||
3ba08129 NH |
358 | /* |
359 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
360 | * on behalf of the thread group. Return task_struct of the (first found) | |
361 | * dedicated thread if found, and return NULL otherwise. | |
362 | * | |
363 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
364 | * have to call rcu_read_lock/unlock() in this function. | |
365 | */ | |
366 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 367 | { |
3ba08129 NH |
368 | struct task_struct *t; |
369 | ||
370 | for_each_thread(tsk, t) | |
371 | if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY)) | |
372 | return t; | |
373 | return NULL; | |
374 | } | |
375 | ||
376 | /* | |
377 | * Determine whether a given process is "early kill" process which expects | |
378 | * to be signaled when some page under the process is hwpoisoned. | |
379 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
380 | * specified) if the process is "early kill," and otherwise returns NULL. | |
381 | */ | |
382 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
383 | int force_early) | |
384 | { | |
385 | struct task_struct *t; | |
6a46079c | 386 | if (!tsk->mm) |
3ba08129 | 387 | return NULL; |
74614de1 | 388 | if (force_early) |
3ba08129 NH |
389 | return tsk; |
390 | t = find_early_kill_thread(tsk); | |
391 | if (t) | |
392 | return t; | |
393 | if (sysctl_memory_failure_early_kill) | |
394 | return tsk; | |
395 | return NULL; | |
6a46079c AK |
396 | } |
397 | ||
398 | /* | |
399 | * Collect processes when the error hit an anonymous page. | |
400 | */ | |
401 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
74614de1 | 402 | struct to_kill **tkc, int force_early) |
6a46079c AK |
403 | { |
404 | struct vm_area_struct *vma; | |
405 | struct task_struct *tsk; | |
406 | struct anon_vma *av; | |
bf181b9f | 407 | pgoff_t pgoff; |
6a46079c | 408 | |
4fc3f1d6 | 409 | av = page_lock_anon_vma_read(page); |
6a46079c | 410 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
411 | return; |
412 | ||
a0f7a756 | 413 | pgoff = page_to_pgoff(page); |
9b679320 | 414 | read_lock(&tasklist_lock); |
6a46079c | 415 | for_each_process (tsk) { |
5beb4930 | 416 | struct anon_vma_chain *vmac; |
3ba08129 | 417 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 418 | |
3ba08129 | 419 | if (!t) |
6a46079c | 420 | continue; |
bf181b9f ML |
421 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
422 | pgoff, pgoff) { | |
5beb4930 | 423 | vma = vmac->vma; |
6a46079c AK |
424 | if (!page_mapped_in_vma(page, vma)) |
425 | continue; | |
3ba08129 NH |
426 | if (vma->vm_mm == t->mm) |
427 | add_to_kill(t, page, vma, to_kill, tkc); | |
6a46079c AK |
428 | } |
429 | } | |
6a46079c | 430 | read_unlock(&tasklist_lock); |
4fc3f1d6 | 431 | page_unlock_anon_vma_read(av); |
6a46079c AK |
432 | } |
433 | ||
434 | /* | |
435 | * Collect processes when the error hit a file mapped page. | |
436 | */ | |
437 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
74614de1 | 438 | struct to_kill **tkc, int force_early) |
6a46079c AK |
439 | { |
440 | struct vm_area_struct *vma; | |
441 | struct task_struct *tsk; | |
6a46079c AK |
442 | struct address_space *mapping = page->mapping; |
443 | ||
d28eb9c8 | 444 | i_mmap_lock_read(mapping); |
9b679320 | 445 | read_lock(&tasklist_lock); |
6a46079c | 446 | for_each_process(tsk) { |
a0f7a756 | 447 | pgoff_t pgoff = page_to_pgoff(page); |
3ba08129 | 448 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 449 | |
3ba08129 | 450 | if (!t) |
6a46079c | 451 | continue; |
6b2dbba8 | 452 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
453 | pgoff) { |
454 | /* | |
455 | * Send early kill signal to tasks where a vma covers | |
456 | * the page but the corrupted page is not necessarily | |
457 | * mapped it in its pte. | |
458 | * Assume applications who requested early kill want | |
459 | * to be informed of all such data corruptions. | |
460 | */ | |
3ba08129 NH |
461 | if (vma->vm_mm == t->mm) |
462 | add_to_kill(t, page, vma, to_kill, tkc); | |
6a46079c AK |
463 | } |
464 | } | |
6a46079c | 465 | read_unlock(&tasklist_lock); |
d28eb9c8 | 466 | i_mmap_unlock_read(mapping); |
6a46079c AK |
467 | } |
468 | ||
469 | /* | |
470 | * Collect the processes who have the corrupted page mapped to kill. | |
471 | * This is done in two steps for locking reasons. | |
472 | * First preallocate one tokill structure outside the spin locks, | |
473 | * so that we can kill at least one process reasonably reliable. | |
474 | */ | |
74614de1 TL |
475 | static void collect_procs(struct page *page, struct list_head *tokill, |
476 | int force_early) | |
6a46079c AK |
477 | { |
478 | struct to_kill *tk; | |
479 | ||
480 | if (!page->mapping) | |
481 | return; | |
482 | ||
483 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
484 | if (!tk) | |
485 | return; | |
486 | if (PageAnon(page)) | |
74614de1 | 487 | collect_procs_anon(page, tokill, &tk, force_early); |
6a46079c | 488 | else |
74614de1 | 489 | collect_procs_file(page, tokill, &tk, force_early); |
6a46079c AK |
490 | kfree(tk); |
491 | } | |
492 | ||
6a46079c | 493 | static const char *action_name[] = { |
cc637b17 XX |
494 | [MF_IGNORED] = "Ignored", |
495 | [MF_FAILED] = "Failed", | |
496 | [MF_DELAYED] = "Delayed", | |
497 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
498 | }; |
499 | ||
500 | static const char * const action_page_types[] = { | |
cc637b17 XX |
501 | [MF_MSG_KERNEL] = "reserved kernel page", |
502 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
503 | [MF_MSG_SLAB] = "kernel slab page", | |
504 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
505 | [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", | |
506 | [MF_MSG_HUGE] = "huge page", | |
507 | [MF_MSG_FREE_HUGE] = "free huge page", | |
508 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", | |
509 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
510 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
511 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
512 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
513 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
514 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
515 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
516 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
517 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
518 | [MF_MSG_BUDDY] = "free buddy page", | |
519 | [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", | |
520 | [MF_MSG_UNKNOWN] = "unknown page", | |
64d37a2b NH |
521 | }; |
522 | ||
dc2a1cbf WF |
523 | /* |
524 | * XXX: It is possible that a page is isolated from LRU cache, | |
525 | * and then kept in swap cache or failed to remove from page cache. | |
526 | * The page count will stop it from being freed by unpoison. | |
527 | * Stress tests should be aware of this memory leak problem. | |
528 | */ | |
529 | static int delete_from_lru_cache(struct page *p) | |
530 | { | |
531 | if (!isolate_lru_page(p)) { | |
532 | /* | |
533 | * Clear sensible page flags, so that the buddy system won't | |
534 | * complain when the page is unpoison-and-freed. | |
535 | */ | |
536 | ClearPageActive(p); | |
537 | ClearPageUnevictable(p); | |
538 | /* | |
539 | * drop the page count elevated by isolate_lru_page() | |
540 | */ | |
09cbfeaf | 541 | put_page(p); |
dc2a1cbf WF |
542 | return 0; |
543 | } | |
544 | return -EIO; | |
545 | } | |
546 | ||
6a46079c AK |
547 | /* |
548 | * Error hit kernel page. | |
549 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
550 | * could be more sophisticated. | |
551 | */ | |
552 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c | 553 | { |
cc637b17 | 554 | return MF_IGNORED; |
6a46079c AK |
555 | } |
556 | ||
557 | /* | |
558 | * Page in unknown state. Do nothing. | |
559 | */ | |
560 | static int me_unknown(struct page *p, unsigned long pfn) | |
561 | { | |
1170532b | 562 | pr_err("MCE %#lx: Unknown page state\n", pfn); |
cc637b17 | 563 | return MF_FAILED; |
6a46079c AK |
564 | } |
565 | ||
6a46079c AK |
566 | /* |
567 | * Clean (or cleaned) page cache page. | |
568 | */ | |
569 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
570 | { | |
571 | int err; | |
cc637b17 | 572 | int ret = MF_FAILED; |
6a46079c AK |
573 | struct address_space *mapping; |
574 | ||
dc2a1cbf WF |
575 | delete_from_lru_cache(p); |
576 | ||
6a46079c AK |
577 | /* |
578 | * For anonymous pages we're done the only reference left | |
579 | * should be the one m_f() holds. | |
580 | */ | |
581 | if (PageAnon(p)) | |
cc637b17 | 582 | return MF_RECOVERED; |
6a46079c AK |
583 | |
584 | /* | |
585 | * Now truncate the page in the page cache. This is really | |
586 | * more like a "temporary hole punch" | |
587 | * Don't do this for block devices when someone else | |
588 | * has a reference, because it could be file system metadata | |
589 | * and that's not safe to truncate. | |
590 | */ | |
591 | mapping = page_mapping(p); | |
592 | if (!mapping) { | |
593 | /* | |
594 | * Page has been teared down in the meanwhile | |
595 | */ | |
cc637b17 | 596 | return MF_FAILED; |
6a46079c AK |
597 | } |
598 | ||
599 | /* | |
600 | * Truncation is a bit tricky. Enable it per file system for now. | |
601 | * | |
602 | * Open: to take i_mutex or not for this? Right now we don't. | |
603 | */ | |
604 | if (mapping->a_ops->error_remove_page) { | |
605 | err = mapping->a_ops->error_remove_page(mapping, p); | |
606 | if (err != 0) { | |
1170532b JP |
607 | pr_info("MCE %#lx: Failed to punch page: %d\n", |
608 | pfn, err); | |
6a46079c AK |
609 | } else if (page_has_private(p) && |
610 | !try_to_release_page(p, GFP_NOIO)) { | |
fb46e735 | 611 | pr_info("MCE %#lx: failed to release buffers\n", pfn); |
6a46079c | 612 | } else { |
cc637b17 | 613 | ret = MF_RECOVERED; |
6a46079c AK |
614 | } |
615 | } else { | |
616 | /* | |
617 | * If the file system doesn't support it just invalidate | |
618 | * This fails on dirty or anything with private pages | |
619 | */ | |
620 | if (invalidate_inode_page(p)) | |
cc637b17 | 621 | ret = MF_RECOVERED; |
6a46079c | 622 | else |
1170532b | 623 | pr_info("MCE %#lx: Failed to invalidate\n", pfn); |
6a46079c AK |
624 | } |
625 | return ret; | |
626 | } | |
627 | ||
628 | /* | |
549543df | 629 | * Dirty pagecache page |
6a46079c AK |
630 | * Issues: when the error hit a hole page the error is not properly |
631 | * propagated. | |
632 | */ | |
633 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
634 | { | |
635 | struct address_space *mapping = page_mapping(p); | |
636 | ||
637 | SetPageError(p); | |
638 | /* TBD: print more information about the file. */ | |
639 | if (mapping) { | |
640 | /* | |
641 | * IO error will be reported by write(), fsync(), etc. | |
642 | * who check the mapping. | |
643 | * This way the application knows that something went | |
644 | * wrong with its dirty file data. | |
645 | * | |
646 | * There's one open issue: | |
647 | * | |
648 | * The EIO will be only reported on the next IO | |
649 | * operation and then cleared through the IO map. | |
650 | * Normally Linux has two mechanisms to pass IO error | |
651 | * first through the AS_EIO flag in the address space | |
652 | * and then through the PageError flag in the page. | |
653 | * Since we drop pages on memory failure handling the | |
654 | * only mechanism open to use is through AS_AIO. | |
655 | * | |
656 | * This has the disadvantage that it gets cleared on | |
657 | * the first operation that returns an error, while | |
658 | * the PageError bit is more sticky and only cleared | |
659 | * when the page is reread or dropped. If an | |
660 | * application assumes it will always get error on | |
661 | * fsync, but does other operations on the fd before | |
25985edc | 662 | * and the page is dropped between then the error |
6a46079c AK |
663 | * will not be properly reported. |
664 | * | |
665 | * This can already happen even without hwpoisoned | |
666 | * pages: first on metadata IO errors (which only | |
667 | * report through AS_EIO) or when the page is dropped | |
668 | * at the wrong time. | |
669 | * | |
670 | * So right now we assume that the application DTRT on | |
671 | * the first EIO, but we're not worse than other parts | |
672 | * of the kernel. | |
673 | */ | |
674 | mapping_set_error(mapping, EIO); | |
675 | } | |
676 | ||
677 | return me_pagecache_clean(p, pfn); | |
678 | } | |
679 | ||
680 | /* | |
681 | * Clean and dirty swap cache. | |
682 | * | |
683 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
684 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
685 | * referenced concurrently by 2 types of PTEs: | |
686 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
687 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
688 | * and then | |
689 | * - clear dirty bit to prevent IO | |
690 | * - remove from LRU | |
691 | * - but keep in the swap cache, so that when we return to it on | |
692 | * a later page fault, we know the application is accessing | |
693 | * corrupted data and shall be killed (we installed simple | |
694 | * interception code in do_swap_page to catch it). | |
695 | * | |
696 | * Clean swap cache pages can be directly isolated. A later page fault will | |
697 | * bring in the known good data from disk. | |
698 | */ | |
699 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
700 | { | |
6a46079c AK |
701 | ClearPageDirty(p); |
702 | /* Trigger EIO in shmem: */ | |
703 | ClearPageUptodate(p); | |
704 | ||
dc2a1cbf | 705 | if (!delete_from_lru_cache(p)) |
cc637b17 | 706 | return MF_DELAYED; |
dc2a1cbf | 707 | else |
cc637b17 | 708 | return MF_FAILED; |
6a46079c AK |
709 | } |
710 | ||
711 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
712 | { | |
6a46079c | 713 | delete_from_swap_cache(p); |
e43c3afb | 714 | |
dc2a1cbf | 715 | if (!delete_from_lru_cache(p)) |
cc637b17 | 716 | return MF_RECOVERED; |
dc2a1cbf | 717 | else |
cc637b17 | 718 | return MF_FAILED; |
6a46079c AK |
719 | } |
720 | ||
721 | /* | |
722 | * Huge pages. Needs work. | |
723 | * Issues: | |
93f70f90 NH |
724 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
725 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
726 | */ |
727 | static int me_huge_page(struct page *p, unsigned long pfn) | |
728 | { | |
6de2b1aa | 729 | int res = 0; |
93f70f90 | 730 | struct page *hpage = compound_head(p); |
2491ffee NH |
731 | |
732 | if (!PageHuge(hpage)) | |
733 | return MF_DELAYED; | |
734 | ||
93f70f90 NH |
735 | /* |
736 | * We can safely recover from error on free or reserved (i.e. | |
737 | * not in-use) hugepage by dequeuing it from freelist. | |
738 | * To check whether a hugepage is in-use or not, we can't use | |
739 | * page->lru because it can be used in other hugepage operations, | |
740 | * such as __unmap_hugepage_range() and gather_surplus_pages(). | |
741 | * So instead we use page_mapping() and PageAnon(). | |
742 | * We assume that this function is called with page lock held, | |
743 | * so there is no race between isolation and mapping/unmapping. | |
744 | */ | |
745 | if (!(page_mapping(hpage) || PageAnon(hpage))) { | |
6de2b1aa NH |
746 | res = dequeue_hwpoisoned_huge_page(hpage); |
747 | if (!res) | |
cc637b17 | 748 | return MF_RECOVERED; |
93f70f90 | 749 | } |
cc637b17 | 750 | return MF_DELAYED; |
6a46079c AK |
751 | } |
752 | ||
753 | /* | |
754 | * Various page states we can handle. | |
755 | * | |
756 | * A page state is defined by its current page->flags bits. | |
757 | * The table matches them in order and calls the right handler. | |
758 | * | |
759 | * This is quite tricky because we can access page at any time | |
25985edc | 760 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
761 | * |
762 | * This is not complete. More states could be added. | |
763 | * For any missing state don't attempt recovery. | |
764 | */ | |
765 | ||
766 | #define dirty (1UL << PG_dirty) | |
767 | #define sc (1UL << PG_swapcache) | |
768 | #define unevict (1UL << PG_unevictable) | |
769 | #define mlock (1UL << PG_mlocked) | |
770 | #define writeback (1UL << PG_writeback) | |
771 | #define lru (1UL << PG_lru) | |
772 | #define swapbacked (1UL << PG_swapbacked) | |
773 | #define head (1UL << PG_head) | |
6a46079c | 774 | #define slab (1UL << PG_slab) |
6a46079c AK |
775 | #define reserved (1UL << PG_reserved) |
776 | ||
777 | static struct page_state { | |
778 | unsigned long mask; | |
779 | unsigned long res; | |
cc637b17 | 780 | enum mf_action_page_type type; |
6a46079c AK |
781 | int (*action)(struct page *p, unsigned long pfn); |
782 | } error_states[] = { | |
cc637b17 | 783 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
784 | /* |
785 | * free pages are specially detected outside this table: | |
786 | * PG_buddy pages only make a small fraction of all free pages. | |
787 | */ | |
6a46079c AK |
788 | |
789 | /* | |
790 | * Could in theory check if slab page is free or if we can drop | |
791 | * currently unused objects without touching them. But just | |
792 | * treat it as standard kernel for now. | |
793 | */ | |
cc637b17 | 794 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 795 | |
cc637b17 | 796 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 797 | |
cc637b17 XX |
798 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
799 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 800 | |
cc637b17 XX |
801 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
802 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 803 | |
cc637b17 XX |
804 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
805 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 806 | |
cc637b17 XX |
807 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
808 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
809 | |
810 | /* | |
811 | * Catchall entry: must be at end. | |
812 | */ | |
cc637b17 | 813 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
814 | }; |
815 | ||
2326c467 AK |
816 | #undef dirty |
817 | #undef sc | |
818 | #undef unevict | |
819 | #undef mlock | |
820 | #undef writeback | |
821 | #undef lru | |
822 | #undef swapbacked | |
823 | #undef head | |
2326c467 AK |
824 | #undef slab |
825 | #undef reserved | |
826 | ||
ff604cf6 NH |
827 | /* |
828 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
829 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
830 | */ | |
cc3e2af4 XX |
831 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
832 | enum mf_result result) | |
6a46079c | 833 | { |
97f0b134 XX |
834 | trace_memory_failure_event(pfn, type, result); |
835 | ||
64d37a2b NH |
836 | pr_err("MCE %#lx: recovery action for %s: %s\n", |
837 | pfn, action_page_types[type], action_name[result]); | |
6a46079c AK |
838 | } |
839 | ||
840 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 841 | unsigned long pfn) |
6a46079c AK |
842 | { |
843 | int result; | |
7456b040 | 844 | int count; |
6a46079c AK |
845 | |
846 | result = ps->action(p, pfn); | |
7456b040 | 847 | |
bd1ce5f9 | 848 | count = page_count(p) - 1; |
cc637b17 | 849 | if (ps->action == me_swapcache_dirty && result == MF_DELAYED) |
138ce286 WF |
850 | count--; |
851 | if (count != 0) { | |
1170532b | 852 | pr_err("MCE %#lx: %s still referenced by %d users\n", |
64d37a2b | 853 | pfn, action_page_types[ps->type], count); |
cc637b17 | 854 | result = MF_FAILED; |
138ce286 | 855 | } |
64d37a2b | 856 | action_result(pfn, ps->type, result); |
6a46079c AK |
857 | |
858 | /* Could do more checks here if page looks ok */ | |
859 | /* | |
860 | * Could adjust zone counters here to correct for the missing page. | |
861 | */ | |
862 | ||
cc637b17 | 863 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
864 | } |
865 | ||
ead07f6a NH |
866 | /** |
867 | * get_hwpoison_page() - Get refcount for memory error handling: | |
868 | * @page: raw error page (hit by memory error) | |
869 | * | |
870 | * Return: return 0 if failed to grab the refcount, otherwise true (some | |
871 | * non-zero value.) | |
872 | */ | |
873 | int get_hwpoison_page(struct page *page) | |
874 | { | |
875 | struct page *head = compound_head(page); | |
876 | ||
4e41a30c | 877 | if (!PageHuge(head) && PageTransHuge(head)) { |
98ed2b00 NH |
878 | /* |
879 | * Non anonymous thp exists only in allocation/free time. We | |
880 | * can't handle such a case correctly, so let's give it up. | |
881 | * This should be better than triggering BUG_ON when kernel | |
882 | * tries to touch the "partially handled" page. | |
883 | */ | |
884 | if (!PageAnon(head)) { | |
885 | pr_err("MCE: %#lx: non anonymous thp\n", | |
886 | page_to_pfn(page)); | |
887 | return 0; | |
888 | } | |
ead07f6a NH |
889 | } |
890 | ||
c2e7e00b KK |
891 | if (get_page_unless_zero(head)) { |
892 | if (head == compound_head(page)) | |
893 | return 1; | |
894 | ||
895 | pr_info("MCE: %#lx cannot catch tail\n", page_to_pfn(page)); | |
896 | put_page(head); | |
897 | } | |
898 | ||
899 | return 0; | |
ead07f6a NH |
900 | } |
901 | EXPORT_SYMBOL_GPL(get_hwpoison_page); | |
902 | ||
6a46079c AK |
903 | /* |
904 | * Do all that is necessary to remove user space mappings. Unmap | |
905 | * the pages and send SIGBUS to the processes if the data was dirty. | |
906 | */ | |
1668bfd5 | 907 | static int hwpoison_user_mappings(struct page *p, unsigned long pfn, |
54b9dd14 | 908 | int trapno, int flags, struct page **hpagep) |
6a46079c AK |
909 | { |
910 | enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
911 | struct address_space *mapping; | |
912 | LIST_HEAD(tokill); | |
913 | int ret; | |
6751ed65 | 914 | int kill = 1, forcekill; |
54b9dd14 | 915 | struct page *hpage = *hpagep; |
6a46079c | 916 | |
93a9eb39 NH |
917 | /* |
918 | * Here we are interested only in user-mapped pages, so skip any | |
919 | * other types of pages. | |
920 | */ | |
921 | if (PageReserved(p) || PageSlab(p)) | |
922 | return SWAP_SUCCESS; | |
923 | if (!(PageLRU(hpage) || PageHuge(p))) | |
1668bfd5 | 924 | return SWAP_SUCCESS; |
6a46079c | 925 | |
6a46079c AK |
926 | /* |
927 | * This check implies we don't kill processes if their pages | |
928 | * are in the swap cache early. Those are always late kills. | |
929 | */ | |
7af446a8 | 930 | if (!page_mapped(hpage)) |
1668bfd5 WF |
931 | return SWAP_SUCCESS; |
932 | ||
52089b14 NH |
933 | if (PageKsm(p)) { |
934 | pr_err("MCE %#lx: can't handle KSM pages.\n", pfn); | |
1668bfd5 | 935 | return SWAP_FAIL; |
52089b14 | 936 | } |
6a46079c AK |
937 | |
938 | if (PageSwapCache(p)) { | |
1170532b | 939 | pr_err("MCE %#lx: keeping poisoned page in swap cache\n", pfn); |
6a46079c AK |
940 | ttu |= TTU_IGNORE_HWPOISON; |
941 | } | |
942 | ||
943 | /* | |
944 | * Propagate the dirty bit from PTEs to struct page first, because we | |
945 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
946 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
947 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 948 | */ |
7af446a8 | 949 | mapping = page_mapping(hpage); |
6751ed65 | 950 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
7af446a8 NH |
951 | mapping_cap_writeback_dirty(mapping)) { |
952 | if (page_mkclean(hpage)) { | |
953 | SetPageDirty(hpage); | |
6a46079c AK |
954 | } else { |
955 | kill = 0; | |
956 | ttu |= TTU_IGNORE_HWPOISON; | |
1170532b | 957 | pr_info("MCE %#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
958 | pfn); |
959 | } | |
960 | } | |
961 | ||
962 | /* | |
963 | * First collect all the processes that have the page | |
964 | * mapped in dirty form. This has to be done before try_to_unmap, | |
965 | * because ttu takes the rmap data structures down. | |
966 | * | |
967 | * Error handling: We ignore errors here because | |
968 | * there's nothing that can be done. | |
969 | */ | |
970 | if (kill) | |
415c64c1 | 971 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 972 | |
415c64c1 | 973 | ret = try_to_unmap(hpage, ttu); |
6a46079c | 974 | if (ret != SWAP_SUCCESS) |
1170532b JP |
975 | pr_err("MCE %#lx: failed to unmap page (mapcount=%d)\n", |
976 | pfn, page_mapcount(hpage)); | |
a6d30ddd | 977 | |
6a46079c AK |
978 | /* |
979 | * Now that the dirty bit has been propagated to the | |
980 | * struct page and all unmaps done we can decide if | |
981 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
982 | * was dirty or the process is not restartable, |
983 | * otherwise the tokill list is merely | |
6a46079c AK |
984 | * freed. When there was a problem unmapping earlier |
985 | * use a more force-full uncatchable kill to prevent | |
986 | * any accesses to the poisoned memory. | |
987 | */ | |
415c64c1 | 988 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
6751ed65 | 989 | kill_procs(&tokill, forcekill, trapno, |
7329bbeb | 990 | ret != SWAP_SUCCESS, p, pfn, flags); |
1668bfd5 WF |
991 | |
992 | return ret; | |
6a46079c AK |
993 | } |
994 | ||
7013febc NH |
995 | static void set_page_hwpoison_huge_page(struct page *hpage) |
996 | { | |
997 | int i; | |
f9121153 | 998 | int nr_pages = 1 << compound_order(hpage); |
7013febc NH |
999 | for (i = 0; i < nr_pages; i++) |
1000 | SetPageHWPoison(hpage + i); | |
1001 | } | |
1002 | ||
1003 | static void clear_page_hwpoison_huge_page(struct page *hpage) | |
1004 | { | |
1005 | int i; | |
f9121153 | 1006 | int nr_pages = 1 << compound_order(hpage); |
7013febc NH |
1007 | for (i = 0; i < nr_pages; i++) |
1008 | ClearPageHWPoison(hpage + i); | |
1009 | } | |
1010 | ||
cd42f4a3 TL |
1011 | /** |
1012 | * memory_failure - Handle memory failure of a page. | |
1013 | * @pfn: Page Number of the corrupted page | |
1014 | * @trapno: Trap number reported in the signal to user space. | |
1015 | * @flags: fine tune action taken | |
1016 | * | |
1017 | * This function is called by the low level machine check code | |
1018 | * of an architecture when it detects hardware memory corruption | |
1019 | * of a page. It tries its best to recover, which includes | |
1020 | * dropping pages, killing processes etc. | |
1021 | * | |
1022 | * The function is primarily of use for corruptions that | |
1023 | * happen outside the current execution context (e.g. when | |
1024 | * detected by a background scrubber) | |
1025 | * | |
1026 | * Must run in process context (e.g. a work queue) with interrupts | |
1027 | * enabled and no spinlocks hold. | |
1028 | */ | |
1029 | int memory_failure(unsigned long pfn, int trapno, int flags) | |
6a46079c AK |
1030 | { |
1031 | struct page_state *ps; | |
1032 | struct page *p; | |
7af446a8 | 1033 | struct page *hpage; |
415c64c1 | 1034 | struct page *orig_head; |
6a46079c | 1035 | int res; |
c9fbdd5f | 1036 | unsigned int nr_pages; |
524fca1e | 1037 | unsigned long page_flags; |
6a46079c AK |
1038 | |
1039 | if (!sysctl_memory_failure_recovery) | |
1040 | panic("Memory failure from trap %d on page %lx", trapno, pfn); | |
1041 | ||
1042 | if (!pfn_valid(pfn)) { | |
1170532b | 1043 | pr_err("MCE %#lx: memory outside kernel control\n", pfn); |
a7560fc8 | 1044 | return -ENXIO; |
6a46079c AK |
1045 | } |
1046 | ||
1047 | p = pfn_to_page(pfn); | |
415c64c1 | 1048 | orig_head = hpage = compound_head(p); |
6a46079c | 1049 | if (TestSetPageHWPoison(p)) { |
1170532b | 1050 | pr_err("MCE %#lx: already hardware poisoned\n", pfn); |
6a46079c AK |
1051 | return 0; |
1052 | } | |
1053 | ||
4db0e950 NH |
1054 | /* |
1055 | * Currently errors on hugetlbfs pages are measured in hugepage units, | |
1056 | * so nr_pages should be 1 << compound_order. OTOH when errors are on | |
1057 | * transparent hugepages, they are supposed to be split and error | |
1058 | * measurement is done in normal page units. So nr_pages should be one | |
1059 | * in this case. | |
1060 | */ | |
1061 | if (PageHuge(p)) | |
1062 | nr_pages = 1 << compound_order(hpage); | |
1063 | else /* normal page or thp */ | |
1064 | nr_pages = 1; | |
8e30456b | 1065 | num_poisoned_pages_add(nr_pages); |
6a46079c AK |
1066 | |
1067 | /* | |
1068 | * We need/can do nothing about count=0 pages. | |
1069 | * 1) it's a free page, and therefore in safe hand: | |
1070 | * prep_new_page() will be the gate keeper. | |
8c6c2ecb NH |
1071 | * 2) it's a free hugepage, which is also safe: |
1072 | * an affected hugepage will be dequeued from hugepage freelist, | |
1073 | * so there's no concern about reusing it ever after. | |
1074 | * 3) it's part of a non-compound high order page. | |
6a46079c AK |
1075 | * Implies some kernel user: cannot stop them from |
1076 | * R/W the page; let's pray that the page has been | |
1077 | * used and will be freed some time later. | |
1078 | * In fact it's dangerous to directly bump up page count from 0, | |
1079 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. | |
1080 | */ | |
ead07f6a | 1081 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { |
8d22ba1b | 1082 | if (is_free_buddy_page(p)) { |
cc637b17 | 1083 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); |
8d22ba1b | 1084 | return 0; |
8c6c2ecb NH |
1085 | } else if (PageHuge(hpage)) { |
1086 | /* | |
b985194c | 1087 | * Check "filter hit" and "race with other subpage." |
8c6c2ecb | 1088 | */ |
7eaceacc | 1089 | lock_page(hpage); |
b985194c CY |
1090 | if (PageHWPoison(hpage)) { |
1091 | if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1092 | || (p != hpage && TestSetPageHWPoison(hpage))) { | |
8e30456b | 1093 | num_poisoned_pages_sub(nr_pages); |
b985194c CY |
1094 | unlock_page(hpage); |
1095 | return 0; | |
1096 | } | |
8c6c2ecb NH |
1097 | } |
1098 | set_page_hwpoison_huge_page(hpage); | |
1099 | res = dequeue_hwpoisoned_huge_page(hpage); | |
cc637b17 XX |
1100 | action_result(pfn, MF_MSG_FREE_HUGE, |
1101 | res ? MF_IGNORED : MF_DELAYED); | |
8c6c2ecb NH |
1102 | unlock_page(hpage); |
1103 | return res; | |
8d22ba1b | 1104 | } else { |
cc637b17 | 1105 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); |
8d22ba1b WF |
1106 | return -EBUSY; |
1107 | } | |
6a46079c AK |
1108 | } |
1109 | ||
415c64c1 | 1110 | if (!PageHuge(p) && PageTransHuge(hpage)) { |
4d2fa965 | 1111 | lock_page(hpage); |
7f6bf39b | 1112 | if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) { |
4d2fa965 | 1113 | unlock_page(hpage); |
7f6bf39b WL |
1114 | if (!PageAnon(hpage)) |
1115 | pr_err("MCE: %#lx: non anonymous thp\n", pfn); | |
1116 | else | |
1117 | pr_err("MCE: %#lx: thp split failed\n", pfn); | |
ead07f6a | 1118 | if (TestClearPageHWPoison(p)) |
8e30456b | 1119 | num_poisoned_pages_sub(nr_pages); |
665d9da7 | 1120 | put_hwpoison_page(p); |
415c64c1 NH |
1121 | return -EBUSY; |
1122 | } | |
4d2fa965 | 1123 | unlock_page(hpage); |
4e41a30c NH |
1124 | get_hwpoison_page(p); |
1125 | put_hwpoison_page(hpage); | |
415c64c1 NH |
1126 | VM_BUG_ON_PAGE(!page_count(p), p); |
1127 | hpage = compound_head(p); | |
1128 | } | |
1129 | ||
e43c3afb WF |
1130 | /* |
1131 | * We ignore non-LRU pages for good reasons. | |
1132 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 1133 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
1134 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
1135 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1136 | * walked by the page reclaim code, however that's not a big loss. | |
1137 | */ | |
09789e5d | 1138 | if (!PageHuge(p)) { |
415c64c1 NH |
1139 | if (!PageLRU(p)) |
1140 | shake_page(p, 0); | |
1141 | if (!PageLRU(p)) { | |
af241a08 JD |
1142 | /* |
1143 | * shake_page could have turned it free. | |
1144 | */ | |
1145 | if (is_free_buddy_page(p)) { | |
2d421acd | 1146 | if (flags & MF_COUNT_INCREASED) |
cc637b17 | 1147 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); |
2d421acd | 1148 | else |
cc637b17 XX |
1149 | action_result(pfn, MF_MSG_BUDDY_2ND, |
1150 | MF_DELAYED); | |
af241a08 JD |
1151 | return 0; |
1152 | } | |
0474a60e | 1153 | } |
e43c3afb | 1154 | } |
e43c3afb | 1155 | |
7eaceacc | 1156 | lock_page(hpage); |
847ce401 | 1157 | |
f37d4298 AK |
1158 | /* |
1159 | * The page could have changed compound pages during the locking. | |
1160 | * If this happens just bail out. | |
1161 | */ | |
415c64c1 | 1162 | if (PageCompound(p) && compound_head(p) != orig_head) { |
cc637b17 | 1163 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 AK |
1164 | res = -EBUSY; |
1165 | goto out; | |
1166 | } | |
1167 | ||
524fca1e NH |
1168 | /* |
1169 | * We use page flags to determine what action should be taken, but | |
1170 | * the flags can be modified by the error containment action. One | |
1171 | * example is an mlocked page, where PG_mlocked is cleared by | |
1172 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1173 | * correctly, we save a copy of the page flags at this time. | |
1174 | */ | |
1175 | page_flags = p->flags; | |
1176 | ||
847ce401 WF |
1177 | /* |
1178 | * unpoison always clear PG_hwpoison inside page lock | |
1179 | */ | |
1180 | if (!PageHWPoison(p)) { | |
1170532b | 1181 | pr_err("MCE %#lx: just unpoisoned\n", pfn); |
8e30456b | 1182 | num_poisoned_pages_sub(nr_pages); |
a09233f3 | 1183 | unlock_page(hpage); |
665d9da7 | 1184 | put_hwpoison_page(hpage); |
a09233f3 | 1185 | return 0; |
847ce401 | 1186 | } |
7c116f2b WF |
1187 | if (hwpoison_filter(p)) { |
1188 | if (TestClearPageHWPoison(p)) | |
8e30456b | 1189 | num_poisoned_pages_sub(nr_pages); |
7af446a8 | 1190 | unlock_page(hpage); |
665d9da7 | 1191 | put_hwpoison_page(hpage); |
7c116f2b WF |
1192 | return 0; |
1193 | } | |
847ce401 | 1194 | |
0bc1f8b0 CY |
1195 | if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p)) |
1196 | goto identify_page_state; | |
1197 | ||
7013febc NH |
1198 | /* |
1199 | * For error on the tail page, we should set PG_hwpoison | |
1200 | * on the head page to show that the hugepage is hwpoisoned | |
1201 | */ | |
a6d30ddd | 1202 | if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) { |
cc637b17 | 1203 | action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED); |
7013febc | 1204 | unlock_page(hpage); |
665d9da7 | 1205 | put_hwpoison_page(hpage); |
7013febc NH |
1206 | return 0; |
1207 | } | |
1208 | /* | |
1209 | * Set PG_hwpoison on all pages in an error hugepage, | |
1210 | * because containment is done in hugepage unit for now. | |
1211 | * Since we have done TestSetPageHWPoison() for the head page with | |
1212 | * page lock held, we can safely set PG_hwpoison bits on tail pages. | |
1213 | */ | |
1214 | if (PageHuge(p)) | |
1215 | set_page_hwpoison_huge_page(hpage); | |
1216 | ||
6edd6cc6 NH |
1217 | /* |
1218 | * It's very difficult to mess with pages currently under IO | |
1219 | * and in many cases impossible, so we just avoid it here. | |
1220 | */ | |
6a46079c AK |
1221 | wait_on_page_writeback(p); |
1222 | ||
1223 | /* | |
1224 | * Now take care of user space mappings. | |
e64a782f | 1225 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
54b9dd14 NH |
1226 | * |
1227 | * When the raw error page is thp tail page, hpage points to the raw | |
1228 | * page after thp split. | |
6a46079c | 1229 | */ |
54b9dd14 NH |
1230 | if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage) |
1231 | != SWAP_SUCCESS) { | |
cc637b17 | 1232 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 WF |
1233 | res = -EBUSY; |
1234 | goto out; | |
1235 | } | |
6a46079c AK |
1236 | |
1237 | /* | |
1238 | * Torn down by someone else? | |
1239 | */ | |
dc2a1cbf | 1240 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 1241 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 1242 | res = -EBUSY; |
6a46079c AK |
1243 | goto out; |
1244 | } | |
1245 | ||
0bc1f8b0 | 1246 | identify_page_state: |
6a46079c | 1247 | res = -EBUSY; |
524fca1e NH |
1248 | /* |
1249 | * The first check uses the current page flags which may not have any | |
1250 | * relevant information. The second check with the saved page flagss is | |
1251 | * carried out only if the first check can't determine the page status. | |
1252 | */ | |
1253 | for (ps = error_states;; ps++) | |
1254 | if ((p->flags & ps->mask) == ps->res) | |
6a46079c | 1255 | break; |
841fcc58 WL |
1256 | |
1257 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1258 | ||
524fca1e NH |
1259 | if (!ps->mask) |
1260 | for (ps = error_states;; ps++) | |
1261 | if ((page_flags & ps->mask) == ps->res) | |
1262 | break; | |
1263 | res = page_action(ps, p, pfn); | |
6a46079c | 1264 | out: |
7af446a8 | 1265 | unlock_page(hpage); |
6a46079c AK |
1266 | return res; |
1267 | } | |
cd42f4a3 | 1268 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1269 | |
ea8f5fb8 HY |
1270 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1271 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1272 | ||
1273 | struct memory_failure_entry { | |
1274 | unsigned long pfn; | |
1275 | int trapno; | |
1276 | int flags; | |
1277 | }; | |
1278 | ||
1279 | struct memory_failure_cpu { | |
1280 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1281 | MEMORY_FAILURE_FIFO_SIZE); | |
1282 | spinlock_t lock; | |
1283 | struct work_struct work; | |
1284 | }; | |
1285 | ||
1286 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1287 | ||
1288 | /** | |
1289 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1290 | * @pfn: Page Number of the corrupted page | |
1291 | * @trapno: Trap number reported in the signal to user space. | |
1292 | * @flags: Flags for memory failure handling | |
1293 | * | |
1294 | * This function is called by the low level hardware error handler | |
1295 | * when it detects hardware memory corruption of a page. It schedules | |
1296 | * the recovering of error page, including dropping pages, killing | |
1297 | * processes etc. | |
1298 | * | |
1299 | * The function is primarily of use for corruptions that | |
1300 | * happen outside the current execution context (e.g. when | |
1301 | * detected by a background scrubber) | |
1302 | * | |
1303 | * Can run in IRQ context. | |
1304 | */ | |
1305 | void memory_failure_queue(unsigned long pfn, int trapno, int flags) | |
1306 | { | |
1307 | struct memory_failure_cpu *mf_cpu; | |
1308 | unsigned long proc_flags; | |
1309 | struct memory_failure_entry entry = { | |
1310 | .pfn = pfn, | |
1311 | .trapno = trapno, | |
1312 | .flags = flags, | |
1313 | }; | |
1314 | ||
1315 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1316 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 1317 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
1318 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
1319 | else | |
8e33a52f | 1320 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
1321 | pfn); |
1322 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1323 | put_cpu_var(memory_failure_cpu); | |
1324 | } | |
1325 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1326 | ||
1327 | static void memory_failure_work_func(struct work_struct *work) | |
1328 | { | |
1329 | struct memory_failure_cpu *mf_cpu; | |
1330 | struct memory_failure_entry entry = { 0, }; | |
1331 | unsigned long proc_flags; | |
1332 | int gotten; | |
1333 | ||
7c8e0181 | 1334 | mf_cpu = this_cpu_ptr(&memory_failure_cpu); |
ea8f5fb8 HY |
1335 | for (;;) { |
1336 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1337 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1338 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1339 | if (!gotten) | |
1340 | break; | |
cf870c70 NR |
1341 | if (entry.flags & MF_SOFT_OFFLINE) |
1342 | soft_offline_page(pfn_to_page(entry.pfn), entry.flags); | |
1343 | else | |
1344 | memory_failure(entry.pfn, entry.trapno, entry.flags); | |
ea8f5fb8 HY |
1345 | } |
1346 | } | |
1347 | ||
1348 | static int __init memory_failure_init(void) | |
1349 | { | |
1350 | struct memory_failure_cpu *mf_cpu; | |
1351 | int cpu; | |
1352 | ||
1353 | for_each_possible_cpu(cpu) { | |
1354 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1355 | spin_lock_init(&mf_cpu->lock); | |
1356 | INIT_KFIFO(mf_cpu->fifo); | |
1357 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1358 | } | |
1359 | ||
1360 | return 0; | |
1361 | } | |
1362 | core_initcall(memory_failure_init); | |
1363 | ||
a5f65109 NH |
1364 | #define unpoison_pr_info(fmt, pfn, rs) \ |
1365 | ({ \ | |
1366 | if (__ratelimit(rs)) \ | |
1367 | pr_info(fmt, pfn); \ | |
1368 | }) | |
1369 | ||
847ce401 WF |
1370 | /** |
1371 | * unpoison_memory - Unpoison a previously poisoned page | |
1372 | * @pfn: Page number of the to be unpoisoned page | |
1373 | * | |
1374 | * Software-unpoison a page that has been poisoned by | |
1375 | * memory_failure() earlier. | |
1376 | * | |
1377 | * This is only done on the software-level, so it only works | |
1378 | * for linux injected failures, not real hardware failures | |
1379 | * | |
1380 | * Returns 0 for success, otherwise -errno. | |
1381 | */ | |
1382 | int unpoison_memory(unsigned long pfn) | |
1383 | { | |
1384 | struct page *page; | |
1385 | struct page *p; | |
1386 | int freeit = 0; | |
c9fbdd5f | 1387 | unsigned int nr_pages; |
a5f65109 NH |
1388 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
1389 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
1390 | |
1391 | if (!pfn_valid(pfn)) | |
1392 | return -ENXIO; | |
1393 | ||
1394 | p = pfn_to_page(pfn); | |
1395 | page = compound_head(p); | |
1396 | ||
1397 | if (!PageHWPoison(p)) { | |
a5f65109 NH |
1398 | unpoison_pr_info("MCE: Page was already unpoisoned %#lx\n", |
1399 | pfn, &unpoison_rs); | |
847ce401 WF |
1400 | return 0; |
1401 | } | |
1402 | ||
230ac719 | 1403 | if (page_count(page) > 1) { |
a5f65109 NH |
1404 | unpoison_pr_info("MCE: Someone grabs the hwpoison page %#lx\n", |
1405 | pfn, &unpoison_rs); | |
230ac719 NH |
1406 | return 0; |
1407 | } | |
1408 | ||
1409 | if (page_mapped(page)) { | |
a5f65109 NH |
1410 | unpoison_pr_info("MCE: Someone maps the hwpoison page %#lx\n", |
1411 | pfn, &unpoison_rs); | |
230ac719 NH |
1412 | return 0; |
1413 | } | |
1414 | ||
1415 | if (page_mapping(page)) { | |
a5f65109 NH |
1416 | unpoison_pr_info("MCE: the hwpoison page has non-NULL mapping %#lx\n", |
1417 | pfn, &unpoison_rs); | |
230ac719 NH |
1418 | return 0; |
1419 | } | |
1420 | ||
0cea3fdc WL |
1421 | /* |
1422 | * unpoison_memory() can encounter thp only when the thp is being | |
1423 | * worked by memory_failure() and the page lock is not held yet. | |
1424 | * In such case, we yield to memory_failure() and make unpoison fail. | |
1425 | */ | |
e76d30e2 | 1426 | if (!PageHuge(page) && PageTransHuge(page)) { |
a5f65109 NH |
1427 | unpoison_pr_info("MCE: Memory failure is now running on %#lx\n", |
1428 | pfn, &unpoison_rs); | |
ead07f6a | 1429 | return 0; |
0cea3fdc WL |
1430 | } |
1431 | ||
f9121153 | 1432 | nr_pages = 1 << compound_order(page); |
c9fbdd5f | 1433 | |
ead07f6a | 1434 | if (!get_hwpoison_page(p)) { |
8c6c2ecb NH |
1435 | /* |
1436 | * Since HWPoisoned hugepage should have non-zero refcount, | |
1437 | * race between memory failure and unpoison seems to happen. | |
1438 | * In such case unpoison fails and memory failure runs | |
1439 | * to the end. | |
1440 | */ | |
1441 | if (PageHuge(page)) { | |
a5f65109 NH |
1442 | unpoison_pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", |
1443 | pfn, &unpoison_rs); | |
8c6c2ecb NH |
1444 | return 0; |
1445 | } | |
847ce401 | 1446 | if (TestClearPageHWPoison(p)) |
8e30456b | 1447 | num_poisoned_pages_dec(); |
a5f65109 NH |
1448 | unpoison_pr_info("MCE: Software-unpoisoned free page %#lx\n", |
1449 | pfn, &unpoison_rs); | |
847ce401 WF |
1450 | return 0; |
1451 | } | |
1452 | ||
7eaceacc | 1453 | lock_page(page); |
847ce401 WF |
1454 | /* |
1455 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1456 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1457 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1458 | * the free buddy page pool. | |
1459 | */ | |
c9fbdd5f | 1460 | if (TestClearPageHWPoison(page)) { |
a5f65109 NH |
1461 | unpoison_pr_info("MCE: Software-unpoisoned page %#lx\n", |
1462 | pfn, &unpoison_rs); | |
8e30456b | 1463 | num_poisoned_pages_sub(nr_pages); |
847ce401 | 1464 | freeit = 1; |
6a90181c NH |
1465 | if (PageHuge(page)) |
1466 | clear_page_hwpoison_huge_page(page); | |
847ce401 WF |
1467 | } |
1468 | unlock_page(page); | |
1469 | ||
665d9da7 | 1470 | put_hwpoison_page(page); |
3ba5eebc | 1471 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) |
665d9da7 | 1472 | put_hwpoison_page(page); |
847ce401 WF |
1473 | |
1474 | return 0; | |
1475 | } | |
1476 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 AK |
1477 | |
1478 | static struct page *new_page(struct page *p, unsigned long private, int **x) | |
1479 | { | |
12686d15 | 1480 | int nid = page_to_nid(p); |
d950b958 NH |
1481 | if (PageHuge(p)) |
1482 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1483 | nid); | |
1484 | else | |
96db800f | 1485 | return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0); |
facb6011 AK |
1486 | } |
1487 | ||
1488 | /* | |
1489 | * Safely get reference count of an arbitrary page. | |
1490 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1491 | * that is not free, and 1 for any other page type. | |
1492 | * For 1 the page is returned with increased page count, otherwise not. | |
1493 | */ | |
af8fae7c | 1494 | static int __get_any_page(struct page *p, unsigned long pfn, int flags) |
facb6011 AK |
1495 | { |
1496 | int ret; | |
1497 | ||
1498 | if (flags & MF_COUNT_INCREASED) | |
1499 | return 1; | |
1500 | ||
d950b958 NH |
1501 | /* |
1502 | * When the target page is a free hugepage, just remove it | |
1503 | * from free hugepage list. | |
1504 | */ | |
ead07f6a | 1505 | if (!get_hwpoison_page(p)) { |
d950b958 | 1506 | if (PageHuge(p)) { |
71dd0b8a | 1507 | pr_info("%s: %#lx free huge page\n", __func__, pfn); |
af8fae7c | 1508 | ret = 0; |
d950b958 | 1509 | } else if (is_free_buddy_page(p)) { |
71dd0b8a | 1510 | pr_info("%s: %#lx free buddy page\n", __func__, pfn); |
facb6011 AK |
1511 | ret = 0; |
1512 | } else { | |
71dd0b8a BP |
1513 | pr_info("%s: %#lx: unknown zero refcount page type %lx\n", |
1514 | __func__, pfn, p->flags); | |
facb6011 AK |
1515 | ret = -EIO; |
1516 | } | |
1517 | } else { | |
1518 | /* Not a free page */ | |
1519 | ret = 1; | |
1520 | } | |
facb6011 AK |
1521 | return ret; |
1522 | } | |
1523 | ||
af8fae7c NH |
1524 | static int get_any_page(struct page *page, unsigned long pfn, int flags) |
1525 | { | |
1526 | int ret = __get_any_page(page, pfn, flags); | |
1527 | ||
1528 | if (ret == 1 && !PageHuge(page) && !PageLRU(page)) { | |
1529 | /* | |
1530 | * Try to free it. | |
1531 | */ | |
665d9da7 | 1532 | put_hwpoison_page(page); |
af8fae7c NH |
1533 | shake_page(page, 1); |
1534 | ||
1535 | /* | |
1536 | * Did it turn free? | |
1537 | */ | |
1538 | ret = __get_any_page(page, pfn, 0); | |
d96b339f | 1539 | if (ret == 1 && !PageLRU(page)) { |
4f32be67 | 1540 | /* Drop page reference which is from __get_any_page() */ |
665d9da7 | 1541 | put_hwpoison_page(page); |
af8fae7c NH |
1542 | pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n", |
1543 | pfn, page->flags); | |
1544 | return -EIO; | |
1545 | } | |
1546 | } | |
1547 | return ret; | |
1548 | } | |
1549 | ||
d950b958 NH |
1550 | static int soft_offline_huge_page(struct page *page, int flags) |
1551 | { | |
1552 | int ret; | |
1553 | unsigned long pfn = page_to_pfn(page); | |
1554 | struct page *hpage = compound_head(page); | |
b8ec1cee | 1555 | LIST_HEAD(pagelist); |
d950b958 | 1556 | |
af8fae7c NH |
1557 | /* |
1558 | * This double-check of PageHWPoison is to avoid the race with | |
1559 | * memory_failure(). See also comment in __soft_offline_page(). | |
1560 | */ | |
1561 | lock_page(hpage); | |
0ebff32c | 1562 | if (PageHWPoison(hpage)) { |
af8fae7c | 1563 | unlock_page(hpage); |
665d9da7 | 1564 | put_hwpoison_page(hpage); |
0ebff32c | 1565 | pr_info("soft offline: %#lx hugepage already poisoned\n", pfn); |
af8fae7c | 1566 | return -EBUSY; |
0ebff32c | 1567 | } |
af8fae7c | 1568 | unlock_page(hpage); |
d950b958 | 1569 | |
bcc54222 | 1570 | ret = isolate_huge_page(hpage, &pagelist); |
03613808 WL |
1571 | /* |
1572 | * get_any_page() and isolate_huge_page() takes a refcount each, | |
1573 | * so need to drop one here. | |
1574 | */ | |
665d9da7 | 1575 | put_hwpoison_page(hpage); |
03613808 | 1576 | if (!ret) { |
bcc54222 NH |
1577 | pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn); |
1578 | return -EBUSY; | |
1579 | } | |
1580 | ||
68711a74 | 1581 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, |
b8ec1cee | 1582 | MIGRATE_SYNC, MR_MEMORY_FAILURE); |
d950b958 | 1583 | if (ret) { |
dd73e85f DN |
1584 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
1585 | pfn, ret, page->flags); | |
b8ec1cee NH |
1586 | /* |
1587 | * We know that soft_offline_huge_page() tries to migrate | |
1588 | * only one hugepage pointed to by hpage, so we need not | |
1589 | * run through the pagelist here. | |
1590 | */ | |
1591 | putback_active_hugepage(hpage); | |
1592 | if (ret > 0) | |
1593 | ret = -EIO; | |
af8fae7c | 1594 | } else { |
a49ecbcd JW |
1595 | /* overcommit hugetlb page will be freed to buddy */ |
1596 | if (PageHuge(page)) { | |
1597 | set_page_hwpoison_huge_page(hpage); | |
1598 | dequeue_hwpoisoned_huge_page(hpage); | |
8e30456b | 1599 | num_poisoned_pages_add(1 << compound_order(hpage)); |
a49ecbcd JW |
1600 | } else { |
1601 | SetPageHWPoison(page); | |
8e30456b | 1602 | num_poisoned_pages_inc(); |
a49ecbcd | 1603 | } |
d950b958 | 1604 | } |
d950b958 NH |
1605 | return ret; |
1606 | } | |
1607 | ||
af8fae7c NH |
1608 | static int __soft_offline_page(struct page *page, int flags) |
1609 | { | |
1610 | int ret; | |
1611 | unsigned long pfn = page_to_pfn(page); | |
facb6011 | 1612 | |
facb6011 | 1613 | /* |
af8fae7c NH |
1614 | * Check PageHWPoison again inside page lock because PageHWPoison |
1615 | * is set by memory_failure() outside page lock. Note that | |
1616 | * memory_failure() also double-checks PageHWPoison inside page lock, | |
1617 | * so there's no race between soft_offline_page() and memory_failure(). | |
facb6011 | 1618 | */ |
0ebff32c XQ |
1619 | lock_page(page); |
1620 | wait_on_page_writeback(page); | |
af8fae7c NH |
1621 | if (PageHWPoison(page)) { |
1622 | unlock_page(page); | |
665d9da7 | 1623 | put_hwpoison_page(page); |
af8fae7c NH |
1624 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
1625 | return -EBUSY; | |
1626 | } | |
facb6011 AK |
1627 | /* |
1628 | * Try to invalidate first. This should work for | |
1629 | * non dirty unmapped page cache pages. | |
1630 | */ | |
1631 | ret = invalidate_inode_page(page); | |
1632 | unlock_page(page); | |
facb6011 | 1633 | /* |
facb6011 AK |
1634 | * RED-PEN would be better to keep it isolated here, but we |
1635 | * would need to fix isolation locking first. | |
1636 | */ | |
facb6011 | 1637 | if (ret == 1) { |
665d9da7 | 1638 | put_hwpoison_page(page); |
fb46e735 | 1639 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
af8fae7c | 1640 | SetPageHWPoison(page); |
8e30456b | 1641 | num_poisoned_pages_inc(); |
af8fae7c | 1642 | return 0; |
facb6011 AK |
1643 | } |
1644 | ||
1645 | /* | |
1646 | * Simple invalidation didn't work. | |
1647 | * Try to migrate to a new page instead. migrate.c | |
1648 | * handles a large number of cases for us. | |
1649 | */ | |
1650 | ret = isolate_lru_page(page); | |
bd486285 KK |
1651 | /* |
1652 | * Drop page reference which is came from get_any_page() | |
1653 | * successful isolate_lru_page() already took another one. | |
1654 | */ | |
665d9da7 | 1655 | put_hwpoison_page(page); |
facb6011 AK |
1656 | if (!ret) { |
1657 | LIST_HEAD(pagelist); | |
5db8a73a | 1658 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
9c620e2b | 1659 | page_is_file_cache(page)); |
facb6011 | 1660 | list_add(&page->lru, &pagelist); |
68711a74 | 1661 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, |
9c620e2b | 1662 | MIGRATE_SYNC, MR_MEMORY_FAILURE); |
facb6011 | 1663 | if (ret) { |
59c82b70 JK |
1664 | if (!list_empty(&pagelist)) { |
1665 | list_del(&page->lru); | |
1666 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
1667 | page_is_file_cache(page)); | |
1668 | putback_lru_page(page); | |
1669 | } | |
1670 | ||
fb46e735 | 1671 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
facb6011 AK |
1672 | pfn, ret, page->flags); |
1673 | if (ret > 0) | |
1674 | ret = -EIO; | |
1675 | } | |
1676 | } else { | |
fb46e735 | 1677 | pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n", |
dd73e85f | 1678 | pfn, ret, page_count(page), page->flags); |
facb6011 | 1679 | } |
facb6011 AK |
1680 | return ret; |
1681 | } | |
86e05773 | 1682 | |
acc14dc4 NH |
1683 | static int soft_offline_in_use_page(struct page *page, int flags) |
1684 | { | |
1685 | int ret; | |
1686 | struct page *hpage = compound_head(page); | |
1687 | ||
1688 | if (!PageHuge(page) && PageTransHuge(hpage)) { | |
1689 | lock_page(hpage); | |
98fd1ef4 NH |
1690 | if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) { |
1691 | unlock_page(hpage); | |
1692 | if (!PageAnon(hpage)) | |
1693 | pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page)); | |
1694 | else | |
1695 | pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page)); | |
1696 | put_hwpoison_page(hpage); | |
acc14dc4 NH |
1697 | return -EBUSY; |
1698 | } | |
98fd1ef4 | 1699 | unlock_page(hpage); |
acc14dc4 NH |
1700 | get_hwpoison_page(page); |
1701 | put_hwpoison_page(hpage); | |
1702 | } | |
1703 | ||
1704 | if (PageHuge(page)) | |
1705 | ret = soft_offline_huge_page(page, flags); | |
1706 | else | |
1707 | ret = __soft_offline_page(page, flags); | |
1708 | ||
1709 | return ret; | |
1710 | } | |
1711 | ||
1712 | static void soft_offline_free_page(struct page *page) | |
1713 | { | |
1714 | if (PageHuge(page)) { | |
1715 | struct page *hpage = compound_head(page); | |
1716 | ||
1717 | set_page_hwpoison_huge_page(hpage); | |
1718 | if (!dequeue_hwpoisoned_huge_page(hpage)) | |
1719 | num_poisoned_pages_add(1 << compound_order(hpage)); | |
1720 | } else { | |
1721 | if (!TestSetPageHWPoison(page)) | |
1722 | num_poisoned_pages_inc(); | |
1723 | } | |
1724 | } | |
1725 | ||
86e05773 WL |
1726 | /** |
1727 | * soft_offline_page - Soft offline a page. | |
1728 | * @page: page to offline | |
1729 | * @flags: flags. Same as memory_failure(). | |
1730 | * | |
1731 | * Returns 0 on success, otherwise negated errno. | |
1732 | * | |
1733 | * Soft offline a page, by migration or invalidation, | |
1734 | * without killing anything. This is for the case when | |
1735 | * a page is not corrupted yet (so it's still valid to access), | |
1736 | * but has had a number of corrected errors and is better taken | |
1737 | * out. | |
1738 | * | |
1739 | * The actual policy on when to do that is maintained by | |
1740 | * user space. | |
1741 | * | |
1742 | * This should never impact any application or cause data loss, | |
1743 | * however it might take some time. | |
1744 | * | |
1745 | * This is not a 100% solution for all memory, but tries to be | |
1746 | * ``good enough'' for the majority of memory. | |
1747 | */ | |
1748 | int soft_offline_page(struct page *page, int flags) | |
1749 | { | |
1750 | int ret; | |
1751 | unsigned long pfn = page_to_pfn(page); | |
86e05773 WL |
1752 | |
1753 | if (PageHWPoison(page)) { | |
1754 | pr_info("soft offline: %#lx page already poisoned\n", pfn); | |
1e0e635b | 1755 | if (flags & MF_COUNT_INCREASED) |
665d9da7 | 1756 | put_hwpoison_page(page); |
86e05773 WL |
1757 | return -EBUSY; |
1758 | } | |
86e05773 | 1759 | |
bfc8c901 | 1760 | get_online_mems(); |
86e05773 | 1761 | ret = get_any_page(page, pfn, flags); |
bfc8c901 | 1762 | put_online_mems(); |
4e41a30c | 1763 | |
acc14dc4 NH |
1764 | if (ret > 0) |
1765 | ret = soft_offline_in_use_page(page, flags); | |
1766 | else if (ret == 0) | |
1767 | soft_offline_free_page(page); | |
4e41a30c | 1768 | |
86e05773 WL |
1769 | return ret; |
1770 | } |