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> |
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 | |
27df5068 AK |
68 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
69 | ||
1bfe5feb | 70 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
71 | u32 hwpoison_filter_dev_major = ~0U; |
72 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
73 | u64 hwpoison_filter_flags_mask; |
74 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 75 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
76 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
77 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
78 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
79 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
80 | |
81 | static int hwpoison_filter_dev(struct page *p) | |
82 | { | |
83 | struct address_space *mapping; | |
84 | dev_t dev; | |
85 | ||
86 | if (hwpoison_filter_dev_major == ~0U && | |
87 | hwpoison_filter_dev_minor == ~0U) | |
88 | return 0; | |
89 | ||
90 | /* | |
1c80b990 | 91 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
92 | */ |
93 | if (PageSlab(p)) | |
94 | return -EINVAL; | |
95 | ||
96 | mapping = page_mapping(p); | |
97 | if (mapping == NULL || mapping->host == NULL) | |
98 | return -EINVAL; | |
99 | ||
100 | dev = mapping->host->i_sb->s_dev; | |
101 | if (hwpoison_filter_dev_major != ~0U && | |
102 | hwpoison_filter_dev_major != MAJOR(dev)) | |
103 | return -EINVAL; | |
104 | if (hwpoison_filter_dev_minor != ~0U && | |
105 | hwpoison_filter_dev_minor != MINOR(dev)) | |
106 | return -EINVAL; | |
107 | ||
108 | return 0; | |
109 | } | |
110 | ||
478c5ffc WF |
111 | static int hwpoison_filter_flags(struct page *p) |
112 | { | |
113 | if (!hwpoison_filter_flags_mask) | |
114 | return 0; | |
115 | ||
116 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
117 | hwpoison_filter_flags_value) | |
118 | return 0; | |
119 | else | |
120 | return -EINVAL; | |
121 | } | |
122 | ||
4fd466eb AK |
123 | /* |
124 | * This allows stress tests to limit test scope to a collection of tasks | |
125 | * by putting them under some memcg. This prevents killing unrelated/important | |
126 | * processes such as /sbin/init. Note that the target task may share clean | |
127 | * pages with init (eg. libc text), which is harmless. If the target task | |
128 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
129 | * is also included in the memcg. At last, due to race conditions this filter | |
130 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
131 | * a freed page. | |
132 | */ | |
94a59fb3 | 133 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
134 | u64 hwpoison_filter_memcg; |
135 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
136 | static int hwpoison_filter_task(struct page *p) | |
137 | { | |
4fd466eb AK |
138 | if (!hwpoison_filter_memcg) |
139 | return 0; | |
140 | ||
94a59fb3 | 141 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
142 | return -EINVAL; |
143 | ||
144 | return 0; | |
145 | } | |
146 | #else | |
147 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
148 | #endif | |
149 | ||
7c116f2b WF |
150 | int hwpoison_filter(struct page *p) |
151 | { | |
1bfe5feb HL |
152 | if (!hwpoison_filter_enable) |
153 | return 0; | |
154 | ||
7c116f2b WF |
155 | if (hwpoison_filter_dev(p)) |
156 | return -EINVAL; | |
157 | ||
478c5ffc WF |
158 | if (hwpoison_filter_flags(p)) |
159 | return -EINVAL; | |
160 | ||
4fd466eb AK |
161 | if (hwpoison_filter_task(p)) |
162 | return -EINVAL; | |
163 | ||
7c116f2b WF |
164 | return 0; |
165 | } | |
27df5068 AK |
166 | #else |
167 | int hwpoison_filter(struct page *p) | |
168 | { | |
169 | return 0; | |
170 | } | |
171 | #endif | |
172 | ||
7c116f2b WF |
173 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
174 | ||
6a46079c | 175 | /* |
7329bbeb TL |
176 | * Send all the processes who have the page mapped a signal. |
177 | * ``action optional'' if they are not immediately affected by the error | |
178 | * ``action required'' if error happened in current execution context | |
6a46079c | 179 | */ |
7329bbeb TL |
180 | static int kill_proc(struct task_struct *t, unsigned long addr, int trapno, |
181 | unsigned long pfn, struct page *page, int flags) | |
6a46079c AK |
182 | { |
183 | struct siginfo si; | |
184 | int ret; | |
185 | ||
186 | printk(KERN_ERR | |
7329bbeb | 187 | "MCE %#lx: Killing %s:%d due to hardware memory corruption\n", |
6a46079c AK |
188 | pfn, t->comm, t->pid); |
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 AK |
210 | if (ret < 0) |
211 | printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n", | |
212 | t->comm, t->pid, ret); | |
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) { | |
292 | printk(KERN_ERR | |
293 | "MCE: Out of memory while machine check handling\n"); | |
294 | return; | |
295 | } | |
296 | } | |
297 | tk->addr = page_address_in_vma(p, vma); | |
298 | tk->addr_valid = 1; | |
299 | ||
300 | /* | |
301 | * In theory we don't have to kill when the page was | |
302 | * munmaped. But it could be also a mremap. Since that's | |
303 | * likely very rare kill anyways just out of paranoia, but use | |
304 | * a SIGKILL because the error is not contained anymore. | |
305 | */ | |
306 | if (tk->addr == -EFAULT) { | |
fb46e735 | 307 | pr_info("MCE: Unable to find user space address %lx in %s\n", |
6a46079c AK |
308 | page_to_pfn(p), tsk->comm); |
309 | tk->addr_valid = 0; | |
310 | } | |
311 | get_task_struct(tsk); | |
312 | tk->tsk = tsk; | |
313 | list_add_tail(&tk->nd, to_kill); | |
314 | } | |
315 | ||
316 | /* | |
317 | * Kill the processes that have been collected earlier. | |
318 | * | |
319 | * Only do anything when DOIT is set, otherwise just free the list | |
320 | * (this is used for clean pages which do not need killing) | |
321 | * Also when FAIL is set do a force kill because something went | |
322 | * wrong earlier. | |
323 | */ | |
6751ed65 | 324 | static void kill_procs(struct list_head *to_kill, int forcekill, int trapno, |
7329bbeb TL |
325 | int fail, struct page *page, unsigned long pfn, |
326 | int flags) | |
6a46079c AK |
327 | { |
328 | struct to_kill *tk, *next; | |
329 | ||
330 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 331 | if (forcekill) { |
6a46079c | 332 | /* |
af901ca1 | 333 | * In case something went wrong with munmapping |
6a46079c AK |
334 | * make sure the process doesn't catch the |
335 | * signal and then access the memory. Just kill it. | |
6a46079c AK |
336 | */ |
337 | if (fail || tk->addr_valid == 0) { | |
338 | printk(KERN_ERR | |
339 | "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", | |
340 | pfn, tk->tsk->comm, tk->tsk->pid); | |
341 | force_sig(SIGKILL, tk->tsk); | |
342 | } | |
343 | ||
344 | /* | |
345 | * In theory the process could have mapped | |
346 | * something else on the address in-between. We could | |
347 | * check for that, but we need to tell the | |
348 | * process anyways. | |
349 | */ | |
7329bbeb TL |
350 | else if (kill_proc(tk->tsk, tk->addr, trapno, |
351 | pfn, page, flags) < 0) | |
6a46079c AK |
352 | printk(KERN_ERR |
353 | "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n", | |
354 | pfn, tk->tsk->comm, tk->tsk->pid); | |
355 | } | |
356 | put_task_struct(tk->tsk); | |
357 | kfree(tk); | |
358 | } | |
359 | } | |
360 | ||
3ba08129 NH |
361 | /* |
362 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
363 | * on behalf of the thread group. Return task_struct of the (first found) | |
364 | * dedicated thread if found, and return NULL otherwise. | |
365 | * | |
366 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
367 | * have to call rcu_read_lock/unlock() in this function. | |
368 | */ | |
369 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 370 | { |
3ba08129 NH |
371 | struct task_struct *t; |
372 | ||
373 | for_each_thread(tsk, t) | |
374 | if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY)) | |
375 | return t; | |
376 | return NULL; | |
377 | } | |
378 | ||
379 | /* | |
380 | * Determine whether a given process is "early kill" process which expects | |
381 | * to be signaled when some page under the process is hwpoisoned. | |
382 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
383 | * specified) if the process is "early kill," and otherwise returns NULL. | |
384 | */ | |
385 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
386 | int force_early) | |
387 | { | |
388 | struct task_struct *t; | |
6a46079c | 389 | if (!tsk->mm) |
3ba08129 | 390 | return NULL; |
74614de1 | 391 | if (force_early) |
3ba08129 NH |
392 | return tsk; |
393 | t = find_early_kill_thread(tsk); | |
394 | if (t) | |
395 | return t; | |
396 | if (sysctl_memory_failure_early_kill) | |
397 | return tsk; | |
398 | return NULL; | |
6a46079c AK |
399 | } |
400 | ||
401 | /* | |
402 | * Collect processes when the error hit an anonymous page. | |
403 | */ | |
404 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
74614de1 | 405 | struct to_kill **tkc, int force_early) |
6a46079c AK |
406 | { |
407 | struct vm_area_struct *vma; | |
408 | struct task_struct *tsk; | |
409 | struct anon_vma *av; | |
bf181b9f | 410 | pgoff_t pgoff; |
6a46079c | 411 | |
4fc3f1d6 | 412 | av = page_lock_anon_vma_read(page); |
6a46079c | 413 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
414 | return; |
415 | ||
a0f7a756 | 416 | pgoff = page_to_pgoff(page); |
9b679320 | 417 | read_lock(&tasklist_lock); |
6a46079c | 418 | for_each_process (tsk) { |
5beb4930 | 419 | struct anon_vma_chain *vmac; |
3ba08129 | 420 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 421 | |
3ba08129 | 422 | if (!t) |
6a46079c | 423 | continue; |
bf181b9f ML |
424 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
425 | pgoff, pgoff) { | |
5beb4930 | 426 | vma = vmac->vma; |
6a46079c AK |
427 | if (!page_mapped_in_vma(page, vma)) |
428 | continue; | |
3ba08129 NH |
429 | if (vma->vm_mm == t->mm) |
430 | add_to_kill(t, page, vma, to_kill, tkc); | |
6a46079c AK |
431 | } |
432 | } | |
6a46079c | 433 | read_unlock(&tasklist_lock); |
4fc3f1d6 | 434 | page_unlock_anon_vma_read(av); |
6a46079c AK |
435 | } |
436 | ||
437 | /* | |
438 | * Collect processes when the error hit a file mapped page. | |
439 | */ | |
440 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
74614de1 | 441 | struct to_kill **tkc, int force_early) |
6a46079c AK |
442 | { |
443 | struct vm_area_struct *vma; | |
444 | struct task_struct *tsk; | |
6a46079c AK |
445 | struct address_space *mapping = page->mapping; |
446 | ||
d28eb9c8 | 447 | i_mmap_lock_read(mapping); |
9b679320 | 448 | read_lock(&tasklist_lock); |
6a46079c | 449 | for_each_process(tsk) { |
a0f7a756 | 450 | pgoff_t pgoff = page_to_pgoff(page); |
3ba08129 | 451 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 452 | |
3ba08129 | 453 | if (!t) |
6a46079c | 454 | continue; |
6b2dbba8 | 455 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
456 | pgoff) { |
457 | /* | |
458 | * Send early kill signal to tasks where a vma covers | |
459 | * the page but the corrupted page is not necessarily | |
460 | * mapped it in its pte. | |
461 | * Assume applications who requested early kill want | |
462 | * to be informed of all such data corruptions. | |
463 | */ | |
3ba08129 NH |
464 | if (vma->vm_mm == t->mm) |
465 | add_to_kill(t, page, vma, to_kill, tkc); | |
6a46079c AK |
466 | } |
467 | } | |
6a46079c | 468 | read_unlock(&tasklist_lock); |
d28eb9c8 | 469 | i_mmap_unlock_read(mapping); |
6a46079c AK |
470 | } |
471 | ||
472 | /* | |
473 | * Collect the processes who have the corrupted page mapped to kill. | |
474 | * This is done in two steps for locking reasons. | |
475 | * First preallocate one tokill structure outside the spin locks, | |
476 | * so that we can kill at least one process reasonably reliable. | |
477 | */ | |
74614de1 TL |
478 | static void collect_procs(struct page *page, struct list_head *tokill, |
479 | int force_early) | |
6a46079c AK |
480 | { |
481 | struct to_kill *tk; | |
482 | ||
483 | if (!page->mapping) | |
484 | return; | |
485 | ||
486 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
487 | if (!tk) | |
488 | return; | |
489 | if (PageAnon(page)) | |
74614de1 | 490 | collect_procs_anon(page, tokill, &tk, force_early); |
6a46079c | 491 | else |
74614de1 | 492 | collect_procs_file(page, tokill, &tk, force_early); |
6a46079c AK |
493 | kfree(tk); |
494 | } | |
495 | ||
6a46079c | 496 | static const char *action_name[] = { |
cc637b17 XX |
497 | [MF_IGNORED] = "Ignored", |
498 | [MF_FAILED] = "Failed", | |
499 | [MF_DELAYED] = "Delayed", | |
500 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
501 | }; |
502 | ||
503 | static const char * const action_page_types[] = { | |
cc637b17 XX |
504 | [MF_MSG_KERNEL] = "reserved kernel page", |
505 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
506 | [MF_MSG_SLAB] = "kernel slab page", | |
507 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
508 | [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", | |
509 | [MF_MSG_HUGE] = "huge page", | |
510 | [MF_MSG_FREE_HUGE] = "free huge page", | |
511 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", | |
512 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
513 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
514 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
515 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
516 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
517 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
518 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
519 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
520 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
521 | [MF_MSG_BUDDY] = "free buddy page", | |
522 | [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", | |
523 | [MF_MSG_UNKNOWN] = "unknown page", | |
64d37a2b NH |
524 | }; |
525 | ||
dc2a1cbf WF |
526 | /* |
527 | * XXX: It is possible that a page is isolated from LRU cache, | |
528 | * and then kept in swap cache or failed to remove from page cache. | |
529 | * The page count will stop it from being freed by unpoison. | |
530 | * Stress tests should be aware of this memory leak problem. | |
531 | */ | |
532 | static int delete_from_lru_cache(struct page *p) | |
533 | { | |
534 | if (!isolate_lru_page(p)) { | |
535 | /* | |
536 | * Clear sensible page flags, so that the buddy system won't | |
537 | * complain when the page is unpoison-and-freed. | |
538 | */ | |
539 | ClearPageActive(p); | |
540 | ClearPageUnevictable(p); | |
541 | /* | |
542 | * drop the page count elevated by isolate_lru_page() | |
543 | */ | |
544 | page_cache_release(p); | |
545 | return 0; | |
546 | } | |
547 | return -EIO; | |
548 | } | |
549 | ||
6a46079c AK |
550 | /* |
551 | * Error hit kernel page. | |
552 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
553 | * could be more sophisticated. | |
554 | */ | |
555 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c | 556 | { |
cc637b17 | 557 | return MF_IGNORED; |
6a46079c AK |
558 | } |
559 | ||
560 | /* | |
561 | * Page in unknown state. Do nothing. | |
562 | */ | |
563 | static int me_unknown(struct page *p, unsigned long pfn) | |
564 | { | |
565 | printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn); | |
cc637b17 | 566 | return MF_FAILED; |
6a46079c AK |
567 | } |
568 | ||
6a46079c AK |
569 | /* |
570 | * Clean (or cleaned) page cache page. | |
571 | */ | |
572 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
573 | { | |
574 | int err; | |
cc637b17 | 575 | int ret = MF_FAILED; |
6a46079c AK |
576 | struct address_space *mapping; |
577 | ||
dc2a1cbf WF |
578 | delete_from_lru_cache(p); |
579 | ||
6a46079c AK |
580 | /* |
581 | * For anonymous pages we're done the only reference left | |
582 | * should be the one m_f() holds. | |
583 | */ | |
584 | if (PageAnon(p)) | |
cc637b17 | 585 | return MF_RECOVERED; |
6a46079c AK |
586 | |
587 | /* | |
588 | * Now truncate the page in the page cache. This is really | |
589 | * more like a "temporary hole punch" | |
590 | * Don't do this for block devices when someone else | |
591 | * has a reference, because it could be file system metadata | |
592 | * and that's not safe to truncate. | |
593 | */ | |
594 | mapping = page_mapping(p); | |
595 | if (!mapping) { | |
596 | /* | |
597 | * Page has been teared down in the meanwhile | |
598 | */ | |
cc637b17 | 599 | return MF_FAILED; |
6a46079c AK |
600 | } |
601 | ||
602 | /* | |
603 | * Truncation is a bit tricky. Enable it per file system for now. | |
604 | * | |
605 | * Open: to take i_mutex or not for this? Right now we don't. | |
606 | */ | |
607 | if (mapping->a_ops->error_remove_page) { | |
608 | err = mapping->a_ops->error_remove_page(mapping, p); | |
609 | if (err != 0) { | |
610 | printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n", | |
611 | pfn, err); | |
612 | } else if (page_has_private(p) && | |
613 | !try_to_release_page(p, GFP_NOIO)) { | |
fb46e735 | 614 | pr_info("MCE %#lx: failed to release buffers\n", pfn); |
6a46079c | 615 | } else { |
cc637b17 | 616 | ret = MF_RECOVERED; |
6a46079c AK |
617 | } |
618 | } else { | |
619 | /* | |
620 | * If the file system doesn't support it just invalidate | |
621 | * This fails on dirty or anything with private pages | |
622 | */ | |
623 | if (invalidate_inode_page(p)) | |
cc637b17 | 624 | ret = MF_RECOVERED; |
6a46079c AK |
625 | else |
626 | printk(KERN_INFO "MCE %#lx: Failed to invalidate\n", | |
627 | pfn); | |
628 | } | |
629 | return ret; | |
630 | } | |
631 | ||
632 | /* | |
549543df | 633 | * Dirty pagecache page |
6a46079c AK |
634 | * Issues: when the error hit a hole page the error is not properly |
635 | * propagated. | |
636 | */ | |
637 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
638 | { | |
639 | struct address_space *mapping = page_mapping(p); | |
640 | ||
641 | SetPageError(p); | |
642 | /* TBD: print more information about the file. */ | |
643 | if (mapping) { | |
644 | /* | |
645 | * IO error will be reported by write(), fsync(), etc. | |
646 | * who check the mapping. | |
647 | * This way the application knows that something went | |
648 | * wrong with its dirty file data. | |
649 | * | |
650 | * There's one open issue: | |
651 | * | |
652 | * The EIO will be only reported on the next IO | |
653 | * operation and then cleared through the IO map. | |
654 | * Normally Linux has two mechanisms to pass IO error | |
655 | * first through the AS_EIO flag in the address space | |
656 | * and then through the PageError flag in the page. | |
657 | * Since we drop pages on memory failure handling the | |
658 | * only mechanism open to use is through AS_AIO. | |
659 | * | |
660 | * This has the disadvantage that it gets cleared on | |
661 | * the first operation that returns an error, while | |
662 | * the PageError bit is more sticky and only cleared | |
663 | * when the page is reread or dropped. If an | |
664 | * application assumes it will always get error on | |
665 | * fsync, but does other operations on the fd before | |
25985edc | 666 | * and the page is dropped between then the error |
6a46079c AK |
667 | * will not be properly reported. |
668 | * | |
669 | * This can already happen even without hwpoisoned | |
670 | * pages: first on metadata IO errors (which only | |
671 | * report through AS_EIO) or when the page is dropped | |
672 | * at the wrong time. | |
673 | * | |
674 | * So right now we assume that the application DTRT on | |
675 | * the first EIO, but we're not worse than other parts | |
676 | * of the kernel. | |
677 | */ | |
678 | mapping_set_error(mapping, EIO); | |
679 | } | |
680 | ||
681 | return me_pagecache_clean(p, pfn); | |
682 | } | |
683 | ||
684 | /* | |
685 | * Clean and dirty swap cache. | |
686 | * | |
687 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
688 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
689 | * referenced concurrently by 2 types of PTEs: | |
690 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
691 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
692 | * and then | |
693 | * - clear dirty bit to prevent IO | |
694 | * - remove from LRU | |
695 | * - but keep in the swap cache, so that when we return to it on | |
696 | * a later page fault, we know the application is accessing | |
697 | * corrupted data and shall be killed (we installed simple | |
698 | * interception code in do_swap_page to catch it). | |
699 | * | |
700 | * Clean swap cache pages can be directly isolated. A later page fault will | |
701 | * bring in the known good data from disk. | |
702 | */ | |
703 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
704 | { | |
6a46079c AK |
705 | ClearPageDirty(p); |
706 | /* Trigger EIO in shmem: */ | |
707 | ClearPageUptodate(p); | |
708 | ||
dc2a1cbf | 709 | if (!delete_from_lru_cache(p)) |
cc637b17 | 710 | return MF_DELAYED; |
dc2a1cbf | 711 | else |
cc637b17 | 712 | return MF_FAILED; |
6a46079c AK |
713 | } |
714 | ||
715 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
716 | { | |
6a46079c | 717 | delete_from_swap_cache(p); |
e43c3afb | 718 | |
dc2a1cbf | 719 | if (!delete_from_lru_cache(p)) |
cc637b17 | 720 | return MF_RECOVERED; |
dc2a1cbf | 721 | else |
cc637b17 | 722 | return MF_FAILED; |
6a46079c AK |
723 | } |
724 | ||
725 | /* | |
726 | * Huge pages. Needs work. | |
727 | * Issues: | |
93f70f90 NH |
728 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
729 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
730 | */ |
731 | static int me_huge_page(struct page *p, unsigned long pfn) | |
732 | { | |
6de2b1aa | 733 | int res = 0; |
93f70f90 | 734 | struct page *hpage = compound_head(p); |
2491ffee NH |
735 | |
736 | if (!PageHuge(hpage)) | |
737 | return MF_DELAYED; | |
738 | ||
93f70f90 NH |
739 | /* |
740 | * We can safely recover from error on free or reserved (i.e. | |
741 | * not in-use) hugepage by dequeuing it from freelist. | |
742 | * To check whether a hugepage is in-use or not, we can't use | |
743 | * page->lru because it can be used in other hugepage operations, | |
744 | * such as __unmap_hugepage_range() and gather_surplus_pages(). | |
745 | * So instead we use page_mapping() and PageAnon(). | |
746 | * We assume that this function is called with page lock held, | |
747 | * so there is no race between isolation and mapping/unmapping. | |
748 | */ | |
749 | if (!(page_mapping(hpage) || PageAnon(hpage))) { | |
6de2b1aa NH |
750 | res = dequeue_hwpoisoned_huge_page(hpage); |
751 | if (!res) | |
cc637b17 | 752 | return MF_RECOVERED; |
93f70f90 | 753 | } |
cc637b17 | 754 | return MF_DELAYED; |
6a46079c AK |
755 | } |
756 | ||
757 | /* | |
758 | * Various page states we can handle. | |
759 | * | |
760 | * A page state is defined by its current page->flags bits. | |
761 | * The table matches them in order and calls the right handler. | |
762 | * | |
763 | * This is quite tricky because we can access page at any time | |
25985edc | 764 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
765 | * |
766 | * This is not complete. More states could be added. | |
767 | * For any missing state don't attempt recovery. | |
768 | */ | |
769 | ||
770 | #define dirty (1UL << PG_dirty) | |
771 | #define sc (1UL << PG_swapcache) | |
772 | #define unevict (1UL << PG_unevictable) | |
773 | #define mlock (1UL << PG_mlocked) | |
774 | #define writeback (1UL << PG_writeback) | |
775 | #define lru (1UL << PG_lru) | |
776 | #define swapbacked (1UL << PG_swapbacked) | |
777 | #define head (1UL << PG_head) | |
778 | #define tail (1UL << PG_tail) | |
779 | #define compound (1UL << PG_compound) | |
780 | #define slab (1UL << PG_slab) | |
6a46079c AK |
781 | #define reserved (1UL << PG_reserved) |
782 | ||
783 | static struct page_state { | |
784 | unsigned long mask; | |
785 | unsigned long res; | |
cc637b17 | 786 | enum mf_action_page_type type; |
6a46079c AK |
787 | int (*action)(struct page *p, unsigned long pfn); |
788 | } error_states[] = { | |
cc637b17 | 789 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
790 | /* |
791 | * free pages are specially detected outside this table: | |
792 | * PG_buddy pages only make a small fraction of all free pages. | |
793 | */ | |
6a46079c AK |
794 | |
795 | /* | |
796 | * Could in theory check if slab page is free or if we can drop | |
797 | * currently unused objects without touching them. But just | |
798 | * treat it as standard kernel for now. | |
799 | */ | |
cc637b17 | 800 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c AK |
801 | |
802 | #ifdef CONFIG_PAGEFLAGS_EXTENDED | |
cc637b17 XX |
803 | { head, head, MF_MSG_HUGE, me_huge_page }, |
804 | { tail, tail, MF_MSG_HUGE, me_huge_page }, | |
6a46079c | 805 | #else |
cc637b17 | 806 | { compound, compound, MF_MSG_HUGE, me_huge_page }, |
6a46079c AK |
807 | #endif |
808 | ||
cc637b17 XX |
809 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
810 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 811 | |
cc637b17 XX |
812 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
813 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 814 | |
cc637b17 XX |
815 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
816 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 817 | |
cc637b17 XX |
818 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
819 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
820 | |
821 | /* | |
822 | * Catchall entry: must be at end. | |
823 | */ | |
cc637b17 | 824 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
825 | }; |
826 | ||
2326c467 AK |
827 | #undef dirty |
828 | #undef sc | |
829 | #undef unevict | |
830 | #undef mlock | |
831 | #undef writeback | |
832 | #undef lru | |
833 | #undef swapbacked | |
834 | #undef head | |
835 | #undef tail | |
836 | #undef compound | |
837 | #undef slab | |
838 | #undef reserved | |
839 | ||
ff604cf6 NH |
840 | /* |
841 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
842 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
843 | */ | |
cc3e2af4 XX |
844 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
845 | enum mf_result result) | |
6a46079c | 846 | { |
97f0b134 XX |
847 | trace_memory_failure_event(pfn, type, result); |
848 | ||
64d37a2b NH |
849 | pr_err("MCE %#lx: recovery action for %s: %s\n", |
850 | pfn, action_page_types[type], action_name[result]); | |
6a46079c AK |
851 | } |
852 | ||
853 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 854 | unsigned long pfn) |
6a46079c AK |
855 | { |
856 | int result; | |
7456b040 | 857 | int count; |
6a46079c AK |
858 | |
859 | result = ps->action(p, pfn); | |
7456b040 | 860 | |
bd1ce5f9 | 861 | count = page_count(p) - 1; |
cc637b17 | 862 | if (ps->action == me_swapcache_dirty && result == MF_DELAYED) |
138ce286 WF |
863 | count--; |
864 | if (count != 0) { | |
6a46079c | 865 | printk(KERN_ERR |
64d37a2b NH |
866 | "MCE %#lx: %s still referenced by %d users\n", |
867 | pfn, action_page_types[ps->type], count); | |
cc637b17 | 868 | result = MF_FAILED; |
138ce286 | 869 | } |
64d37a2b | 870 | action_result(pfn, ps->type, result); |
6a46079c AK |
871 | |
872 | /* Could do more checks here if page looks ok */ | |
873 | /* | |
874 | * Could adjust zone counters here to correct for the missing page. | |
875 | */ | |
876 | ||
cc637b17 | 877 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
878 | } |
879 | ||
ead07f6a NH |
880 | /** |
881 | * get_hwpoison_page() - Get refcount for memory error handling: | |
882 | * @page: raw error page (hit by memory error) | |
883 | * | |
884 | * Return: return 0 if failed to grab the refcount, otherwise true (some | |
885 | * non-zero value.) | |
886 | */ | |
887 | int get_hwpoison_page(struct page *page) | |
888 | { | |
889 | struct page *head = compound_head(page); | |
890 | ||
891 | if (PageHuge(head)) | |
892 | return get_page_unless_zero(head); | |
893 | ||
894 | /* | |
895 | * Thp tail page has special refcounting rule (refcount of tail pages | |
896 | * is stored in ->_mapcount,) so we can't call get_page_unless_zero() | |
897 | * directly for tail pages. | |
898 | */ | |
899 | if (PageTransHuge(head)) { | |
98ed2b00 NH |
900 | /* |
901 | * Non anonymous thp exists only in allocation/free time. We | |
902 | * can't handle such a case correctly, so let's give it up. | |
903 | * This should be better than triggering BUG_ON when kernel | |
904 | * tries to touch the "partially handled" page. | |
905 | */ | |
906 | if (!PageAnon(head)) { | |
907 | pr_err("MCE: %#lx: non anonymous thp\n", | |
908 | page_to_pfn(page)); | |
909 | return 0; | |
910 | } | |
911 | ||
ead07f6a NH |
912 | if (get_page_unless_zero(head)) { |
913 | if (PageTail(page)) | |
914 | get_page(page); | |
915 | return 1; | |
916 | } else { | |
917 | return 0; | |
918 | } | |
919 | } | |
920 | ||
921 | return get_page_unless_zero(page); | |
922 | } | |
923 | EXPORT_SYMBOL_GPL(get_hwpoison_page); | |
924 | ||
94bf4ec8 WL |
925 | /** |
926 | * put_hwpoison_page() - Put refcount for memory error handling: | |
927 | * @page: raw error page (hit by memory error) | |
928 | */ | |
929 | void put_hwpoison_page(struct page *page) | |
930 | { | |
931 | struct page *head = compound_head(page); | |
932 | ||
933 | if (PageHuge(head)) { | |
934 | put_page(head); | |
935 | return; | |
936 | } | |
937 | ||
938 | if (PageTransHuge(head)) | |
939 | if (page != head) | |
940 | put_page(head); | |
941 | ||
942 | put_page(page); | |
943 | } | |
944 | EXPORT_SYMBOL_GPL(put_hwpoison_page); | |
945 | ||
6a46079c AK |
946 | /* |
947 | * Do all that is necessary to remove user space mappings. Unmap | |
948 | * the pages and send SIGBUS to the processes if the data was dirty. | |
949 | */ | |
1668bfd5 | 950 | static int hwpoison_user_mappings(struct page *p, unsigned long pfn, |
54b9dd14 | 951 | int trapno, int flags, struct page **hpagep) |
6a46079c AK |
952 | { |
953 | enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; | |
954 | struct address_space *mapping; | |
955 | LIST_HEAD(tokill); | |
956 | int ret; | |
6751ed65 | 957 | int kill = 1, forcekill; |
54b9dd14 | 958 | struct page *hpage = *hpagep; |
6a46079c | 959 | |
93a9eb39 NH |
960 | /* |
961 | * Here we are interested only in user-mapped pages, so skip any | |
962 | * other types of pages. | |
963 | */ | |
964 | if (PageReserved(p) || PageSlab(p)) | |
965 | return SWAP_SUCCESS; | |
966 | if (!(PageLRU(hpage) || PageHuge(p))) | |
1668bfd5 | 967 | return SWAP_SUCCESS; |
6a46079c | 968 | |
6a46079c AK |
969 | /* |
970 | * This check implies we don't kill processes if their pages | |
971 | * are in the swap cache early. Those are always late kills. | |
972 | */ | |
7af446a8 | 973 | if (!page_mapped(hpage)) |
1668bfd5 WF |
974 | return SWAP_SUCCESS; |
975 | ||
52089b14 NH |
976 | if (PageKsm(p)) { |
977 | pr_err("MCE %#lx: can't handle KSM pages.\n", pfn); | |
1668bfd5 | 978 | return SWAP_FAIL; |
52089b14 | 979 | } |
6a46079c AK |
980 | |
981 | if (PageSwapCache(p)) { | |
982 | printk(KERN_ERR | |
983 | "MCE %#lx: keeping poisoned page in swap cache\n", pfn); | |
984 | ttu |= TTU_IGNORE_HWPOISON; | |
985 | } | |
986 | ||
987 | /* | |
988 | * Propagate the dirty bit from PTEs to struct page first, because we | |
989 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
990 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
991 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 992 | */ |
7af446a8 | 993 | mapping = page_mapping(hpage); |
6751ed65 | 994 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
7af446a8 NH |
995 | mapping_cap_writeback_dirty(mapping)) { |
996 | if (page_mkclean(hpage)) { | |
997 | SetPageDirty(hpage); | |
6a46079c AK |
998 | } else { |
999 | kill = 0; | |
1000 | ttu |= TTU_IGNORE_HWPOISON; | |
1001 | printk(KERN_INFO | |
1002 | "MCE %#lx: corrupted page was clean: dropped without side effects\n", | |
1003 | pfn); | |
1004 | } | |
1005 | } | |
1006 | ||
1007 | /* | |
1008 | * First collect all the processes that have the page | |
1009 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1010 | * because ttu takes the rmap data structures down. | |
1011 | * | |
1012 | * Error handling: We ignore errors here because | |
1013 | * there's nothing that can be done. | |
1014 | */ | |
1015 | if (kill) | |
415c64c1 | 1016 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1017 | |
415c64c1 | 1018 | ret = try_to_unmap(hpage, ttu); |
6a46079c AK |
1019 | if (ret != SWAP_SUCCESS) |
1020 | printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n", | |
415c64c1 | 1021 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1022 | |
6a46079c AK |
1023 | /* |
1024 | * Now that the dirty bit has been propagated to the | |
1025 | * struct page and all unmaps done we can decide if | |
1026 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1027 | * was dirty or the process is not restartable, |
1028 | * otherwise the tokill list is merely | |
6a46079c AK |
1029 | * freed. When there was a problem unmapping earlier |
1030 | * use a more force-full uncatchable kill to prevent | |
1031 | * any accesses to the poisoned memory. | |
1032 | */ | |
415c64c1 | 1033 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
6751ed65 | 1034 | kill_procs(&tokill, forcekill, trapno, |
7329bbeb | 1035 | ret != SWAP_SUCCESS, p, pfn, flags); |
1668bfd5 WF |
1036 | |
1037 | return ret; | |
6a46079c AK |
1038 | } |
1039 | ||
7013febc NH |
1040 | static void set_page_hwpoison_huge_page(struct page *hpage) |
1041 | { | |
1042 | int i; | |
f9121153 | 1043 | int nr_pages = 1 << compound_order(hpage); |
7013febc NH |
1044 | for (i = 0; i < nr_pages; i++) |
1045 | SetPageHWPoison(hpage + i); | |
1046 | } | |
1047 | ||
1048 | static void clear_page_hwpoison_huge_page(struct page *hpage) | |
1049 | { | |
1050 | int i; | |
f9121153 | 1051 | int nr_pages = 1 << compound_order(hpage); |
7013febc NH |
1052 | for (i = 0; i < nr_pages; i++) |
1053 | ClearPageHWPoison(hpage + i); | |
1054 | } | |
1055 | ||
cd42f4a3 TL |
1056 | /** |
1057 | * memory_failure - Handle memory failure of a page. | |
1058 | * @pfn: Page Number of the corrupted page | |
1059 | * @trapno: Trap number reported in the signal to user space. | |
1060 | * @flags: fine tune action taken | |
1061 | * | |
1062 | * This function is called by the low level machine check code | |
1063 | * of an architecture when it detects hardware memory corruption | |
1064 | * of a page. It tries its best to recover, which includes | |
1065 | * dropping pages, killing processes etc. | |
1066 | * | |
1067 | * The function is primarily of use for corruptions that | |
1068 | * happen outside the current execution context (e.g. when | |
1069 | * detected by a background scrubber) | |
1070 | * | |
1071 | * Must run in process context (e.g. a work queue) with interrupts | |
1072 | * enabled and no spinlocks hold. | |
1073 | */ | |
1074 | int memory_failure(unsigned long pfn, int trapno, int flags) | |
6a46079c AK |
1075 | { |
1076 | struct page_state *ps; | |
1077 | struct page *p; | |
7af446a8 | 1078 | struct page *hpage; |
415c64c1 | 1079 | struct page *orig_head; |
6a46079c | 1080 | int res; |
c9fbdd5f | 1081 | unsigned int nr_pages; |
524fca1e | 1082 | unsigned long page_flags; |
6a46079c AK |
1083 | |
1084 | if (!sysctl_memory_failure_recovery) | |
1085 | panic("Memory failure from trap %d on page %lx", trapno, pfn); | |
1086 | ||
1087 | if (!pfn_valid(pfn)) { | |
a7560fc8 WF |
1088 | printk(KERN_ERR |
1089 | "MCE %#lx: memory outside kernel control\n", | |
1090 | pfn); | |
1091 | return -ENXIO; | |
6a46079c AK |
1092 | } |
1093 | ||
1094 | p = pfn_to_page(pfn); | |
415c64c1 | 1095 | orig_head = hpage = compound_head(p); |
6a46079c | 1096 | if (TestSetPageHWPoison(p)) { |
d95ea51e | 1097 | printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn); |
6a46079c AK |
1098 | return 0; |
1099 | } | |
1100 | ||
4db0e950 NH |
1101 | /* |
1102 | * Currently errors on hugetlbfs pages are measured in hugepage units, | |
1103 | * so nr_pages should be 1 << compound_order. OTOH when errors are on | |
1104 | * transparent hugepages, they are supposed to be split and error | |
1105 | * measurement is done in normal page units. So nr_pages should be one | |
1106 | * in this case. | |
1107 | */ | |
1108 | if (PageHuge(p)) | |
1109 | nr_pages = 1 << compound_order(hpage); | |
1110 | else /* normal page or thp */ | |
1111 | nr_pages = 1; | |
8e30456b | 1112 | num_poisoned_pages_add(nr_pages); |
6a46079c AK |
1113 | |
1114 | /* | |
1115 | * We need/can do nothing about count=0 pages. | |
1116 | * 1) it's a free page, and therefore in safe hand: | |
1117 | * prep_new_page() will be the gate keeper. | |
8c6c2ecb NH |
1118 | * 2) it's a free hugepage, which is also safe: |
1119 | * an affected hugepage will be dequeued from hugepage freelist, | |
1120 | * so there's no concern about reusing it ever after. | |
1121 | * 3) it's part of a non-compound high order page. | |
6a46079c AK |
1122 | * Implies some kernel user: cannot stop them from |
1123 | * R/W the page; let's pray that the page has been | |
1124 | * used and will be freed some time later. | |
1125 | * In fact it's dangerous to directly bump up page count from 0, | |
1126 | * that may make page_freeze_refs()/page_unfreeze_refs() mismatch. | |
1127 | */ | |
ead07f6a | 1128 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { |
8d22ba1b | 1129 | if (is_free_buddy_page(p)) { |
cc637b17 | 1130 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); |
8d22ba1b | 1131 | return 0; |
8c6c2ecb NH |
1132 | } else if (PageHuge(hpage)) { |
1133 | /* | |
b985194c | 1134 | * Check "filter hit" and "race with other subpage." |
8c6c2ecb | 1135 | */ |
7eaceacc | 1136 | lock_page(hpage); |
b985194c CY |
1137 | if (PageHWPoison(hpage)) { |
1138 | if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1139 | || (p != hpage && TestSetPageHWPoison(hpage))) { | |
8e30456b | 1140 | num_poisoned_pages_sub(nr_pages); |
b985194c CY |
1141 | unlock_page(hpage); |
1142 | return 0; | |
1143 | } | |
8c6c2ecb NH |
1144 | } |
1145 | set_page_hwpoison_huge_page(hpage); | |
1146 | res = dequeue_hwpoisoned_huge_page(hpage); | |
cc637b17 XX |
1147 | action_result(pfn, MF_MSG_FREE_HUGE, |
1148 | res ? MF_IGNORED : MF_DELAYED); | |
8c6c2ecb NH |
1149 | unlock_page(hpage); |
1150 | return res; | |
8d22ba1b | 1151 | } else { |
cc637b17 | 1152 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); |
8d22ba1b WF |
1153 | return -EBUSY; |
1154 | } | |
6a46079c AK |
1155 | } |
1156 | ||
415c64c1 | 1157 | if (!PageHuge(p) && PageTransHuge(hpage)) { |
7f6bf39b WL |
1158 | if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) { |
1159 | if (!PageAnon(hpage)) | |
1160 | pr_err("MCE: %#lx: non anonymous thp\n", pfn); | |
1161 | else | |
1162 | pr_err("MCE: %#lx: thp split failed\n", pfn); | |
ead07f6a | 1163 | if (TestClearPageHWPoison(p)) |
8e30456b | 1164 | num_poisoned_pages_sub(nr_pages); |
665d9da7 | 1165 | put_hwpoison_page(p); |
415c64c1 NH |
1166 | return -EBUSY; |
1167 | } | |
1168 | VM_BUG_ON_PAGE(!page_count(p), p); | |
1169 | hpage = compound_head(p); | |
1170 | } | |
1171 | ||
e43c3afb WF |
1172 | /* |
1173 | * We ignore non-LRU pages for good reasons. | |
1174 | * - PG_locked is only well defined for LRU pages and a few others | |
1175 | * - to avoid races with __set_page_locked() | |
1176 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) | |
1177 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1178 | * walked by the page reclaim code, however that's not a big loss. | |
1179 | */ | |
09789e5d | 1180 | if (!PageHuge(p)) { |
415c64c1 NH |
1181 | if (!PageLRU(p)) |
1182 | shake_page(p, 0); | |
1183 | if (!PageLRU(p)) { | |
af241a08 JD |
1184 | /* |
1185 | * shake_page could have turned it free. | |
1186 | */ | |
1187 | if (is_free_buddy_page(p)) { | |
2d421acd | 1188 | if (flags & MF_COUNT_INCREASED) |
cc637b17 | 1189 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); |
2d421acd | 1190 | else |
cc637b17 XX |
1191 | action_result(pfn, MF_MSG_BUDDY_2ND, |
1192 | MF_DELAYED); | |
af241a08 JD |
1193 | return 0; |
1194 | } | |
0474a60e | 1195 | } |
e43c3afb | 1196 | } |
e43c3afb | 1197 | |
7eaceacc | 1198 | lock_page(hpage); |
847ce401 | 1199 | |
f37d4298 AK |
1200 | /* |
1201 | * The page could have changed compound pages during the locking. | |
1202 | * If this happens just bail out. | |
1203 | */ | |
415c64c1 | 1204 | if (PageCompound(p) && compound_head(p) != orig_head) { |
cc637b17 | 1205 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 AK |
1206 | res = -EBUSY; |
1207 | goto out; | |
1208 | } | |
1209 | ||
524fca1e NH |
1210 | /* |
1211 | * We use page flags to determine what action should be taken, but | |
1212 | * the flags can be modified by the error containment action. One | |
1213 | * example is an mlocked page, where PG_mlocked is cleared by | |
1214 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1215 | * correctly, we save a copy of the page flags at this time. | |
1216 | */ | |
1217 | page_flags = p->flags; | |
1218 | ||
847ce401 WF |
1219 | /* |
1220 | * unpoison always clear PG_hwpoison inside page lock | |
1221 | */ | |
1222 | if (!PageHWPoison(p)) { | |
d95ea51e | 1223 | printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn); |
8e30456b | 1224 | num_poisoned_pages_sub(nr_pages); |
a09233f3 | 1225 | unlock_page(hpage); |
665d9da7 | 1226 | put_hwpoison_page(hpage); |
a09233f3 | 1227 | return 0; |
847ce401 | 1228 | } |
7c116f2b WF |
1229 | if (hwpoison_filter(p)) { |
1230 | if (TestClearPageHWPoison(p)) | |
8e30456b | 1231 | num_poisoned_pages_sub(nr_pages); |
7af446a8 | 1232 | unlock_page(hpage); |
665d9da7 | 1233 | put_hwpoison_page(hpage); |
7c116f2b WF |
1234 | return 0; |
1235 | } | |
847ce401 | 1236 | |
0bc1f8b0 CY |
1237 | if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p)) |
1238 | goto identify_page_state; | |
1239 | ||
7013febc NH |
1240 | /* |
1241 | * For error on the tail page, we should set PG_hwpoison | |
1242 | * on the head page to show that the hugepage is hwpoisoned | |
1243 | */ | |
a6d30ddd | 1244 | if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) { |
cc637b17 | 1245 | action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED); |
7013febc | 1246 | unlock_page(hpage); |
665d9da7 | 1247 | put_hwpoison_page(hpage); |
7013febc NH |
1248 | return 0; |
1249 | } | |
1250 | /* | |
1251 | * Set PG_hwpoison on all pages in an error hugepage, | |
1252 | * because containment is done in hugepage unit for now. | |
1253 | * Since we have done TestSetPageHWPoison() for the head page with | |
1254 | * page lock held, we can safely set PG_hwpoison bits on tail pages. | |
1255 | */ | |
1256 | if (PageHuge(p)) | |
1257 | set_page_hwpoison_huge_page(hpage); | |
1258 | ||
6edd6cc6 NH |
1259 | /* |
1260 | * It's very difficult to mess with pages currently under IO | |
1261 | * and in many cases impossible, so we just avoid it here. | |
1262 | */ | |
6a46079c AK |
1263 | wait_on_page_writeback(p); |
1264 | ||
1265 | /* | |
1266 | * Now take care of user space mappings. | |
e64a782f | 1267 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
54b9dd14 NH |
1268 | * |
1269 | * When the raw error page is thp tail page, hpage points to the raw | |
1270 | * page after thp split. | |
6a46079c | 1271 | */ |
54b9dd14 NH |
1272 | if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage) |
1273 | != SWAP_SUCCESS) { | |
cc637b17 | 1274 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 WF |
1275 | res = -EBUSY; |
1276 | goto out; | |
1277 | } | |
6a46079c AK |
1278 | |
1279 | /* | |
1280 | * Torn down by someone else? | |
1281 | */ | |
dc2a1cbf | 1282 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 1283 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 1284 | res = -EBUSY; |
6a46079c AK |
1285 | goto out; |
1286 | } | |
1287 | ||
0bc1f8b0 | 1288 | identify_page_state: |
6a46079c | 1289 | res = -EBUSY; |
524fca1e NH |
1290 | /* |
1291 | * The first check uses the current page flags which may not have any | |
1292 | * relevant information. The second check with the saved page flagss is | |
1293 | * carried out only if the first check can't determine the page status. | |
1294 | */ | |
1295 | for (ps = error_states;; ps++) | |
1296 | if ((p->flags & ps->mask) == ps->res) | |
6a46079c | 1297 | break; |
841fcc58 WL |
1298 | |
1299 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1300 | ||
524fca1e NH |
1301 | if (!ps->mask) |
1302 | for (ps = error_states;; ps++) | |
1303 | if ((page_flags & ps->mask) == ps->res) | |
1304 | break; | |
1305 | res = page_action(ps, p, pfn); | |
6a46079c | 1306 | out: |
7af446a8 | 1307 | unlock_page(hpage); |
6a46079c AK |
1308 | return res; |
1309 | } | |
cd42f4a3 | 1310 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1311 | |
ea8f5fb8 HY |
1312 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1313 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1314 | ||
1315 | struct memory_failure_entry { | |
1316 | unsigned long pfn; | |
1317 | int trapno; | |
1318 | int flags; | |
1319 | }; | |
1320 | ||
1321 | struct memory_failure_cpu { | |
1322 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1323 | MEMORY_FAILURE_FIFO_SIZE); | |
1324 | spinlock_t lock; | |
1325 | struct work_struct work; | |
1326 | }; | |
1327 | ||
1328 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1329 | ||
1330 | /** | |
1331 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1332 | * @pfn: Page Number of the corrupted page | |
1333 | * @trapno: Trap number reported in the signal to user space. | |
1334 | * @flags: Flags for memory failure handling | |
1335 | * | |
1336 | * This function is called by the low level hardware error handler | |
1337 | * when it detects hardware memory corruption of a page. It schedules | |
1338 | * the recovering of error page, including dropping pages, killing | |
1339 | * processes etc. | |
1340 | * | |
1341 | * The function is primarily of use for corruptions that | |
1342 | * happen outside the current execution context (e.g. when | |
1343 | * detected by a background scrubber) | |
1344 | * | |
1345 | * Can run in IRQ context. | |
1346 | */ | |
1347 | void memory_failure_queue(unsigned long pfn, int trapno, int flags) | |
1348 | { | |
1349 | struct memory_failure_cpu *mf_cpu; | |
1350 | unsigned long proc_flags; | |
1351 | struct memory_failure_entry entry = { | |
1352 | .pfn = pfn, | |
1353 | .trapno = trapno, | |
1354 | .flags = flags, | |
1355 | }; | |
1356 | ||
1357 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1358 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 1359 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
1360 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
1361 | else | |
8e33a52f | 1362 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
1363 | pfn); |
1364 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1365 | put_cpu_var(memory_failure_cpu); | |
1366 | } | |
1367 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1368 | ||
1369 | static void memory_failure_work_func(struct work_struct *work) | |
1370 | { | |
1371 | struct memory_failure_cpu *mf_cpu; | |
1372 | struct memory_failure_entry entry = { 0, }; | |
1373 | unsigned long proc_flags; | |
1374 | int gotten; | |
1375 | ||
7c8e0181 | 1376 | mf_cpu = this_cpu_ptr(&memory_failure_cpu); |
ea8f5fb8 HY |
1377 | for (;;) { |
1378 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1379 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1380 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1381 | if (!gotten) | |
1382 | break; | |
cf870c70 NR |
1383 | if (entry.flags & MF_SOFT_OFFLINE) |
1384 | soft_offline_page(pfn_to_page(entry.pfn), entry.flags); | |
1385 | else | |
1386 | memory_failure(entry.pfn, entry.trapno, entry.flags); | |
ea8f5fb8 HY |
1387 | } |
1388 | } | |
1389 | ||
1390 | static int __init memory_failure_init(void) | |
1391 | { | |
1392 | struct memory_failure_cpu *mf_cpu; | |
1393 | int cpu; | |
1394 | ||
1395 | for_each_possible_cpu(cpu) { | |
1396 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1397 | spin_lock_init(&mf_cpu->lock); | |
1398 | INIT_KFIFO(mf_cpu->fifo); | |
1399 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1400 | } | |
1401 | ||
1402 | return 0; | |
1403 | } | |
1404 | core_initcall(memory_failure_init); | |
1405 | ||
847ce401 WF |
1406 | /** |
1407 | * unpoison_memory - Unpoison a previously poisoned page | |
1408 | * @pfn: Page number of the to be unpoisoned page | |
1409 | * | |
1410 | * Software-unpoison a page that has been poisoned by | |
1411 | * memory_failure() earlier. | |
1412 | * | |
1413 | * This is only done on the software-level, so it only works | |
1414 | * for linux injected failures, not real hardware failures | |
1415 | * | |
1416 | * Returns 0 for success, otherwise -errno. | |
1417 | */ | |
1418 | int unpoison_memory(unsigned long pfn) | |
1419 | { | |
1420 | struct page *page; | |
1421 | struct page *p; | |
1422 | int freeit = 0; | |
c9fbdd5f | 1423 | unsigned int nr_pages; |
847ce401 WF |
1424 | |
1425 | if (!pfn_valid(pfn)) | |
1426 | return -ENXIO; | |
1427 | ||
1428 | p = pfn_to_page(pfn); | |
1429 | page = compound_head(p); | |
1430 | ||
1431 | if (!PageHWPoison(p)) { | |
fb46e735 | 1432 | pr_info("MCE: Page was already unpoisoned %#lx\n", pfn); |
847ce401 WF |
1433 | return 0; |
1434 | } | |
1435 | ||
230ac719 NH |
1436 | if (page_count(page) > 1) { |
1437 | pr_info("MCE: Someone grabs the hwpoison page %#lx\n", pfn); | |
1438 | return 0; | |
1439 | } | |
1440 | ||
1441 | if (page_mapped(page)) { | |
1442 | pr_info("MCE: Someone maps the hwpoison page %#lx\n", pfn); | |
1443 | return 0; | |
1444 | } | |
1445 | ||
1446 | if (page_mapping(page)) { | |
1447 | pr_info("MCE: the hwpoison page has non-NULL mapping %#lx\n", | |
1448 | pfn); | |
1449 | return 0; | |
1450 | } | |
1451 | ||
0cea3fdc WL |
1452 | /* |
1453 | * unpoison_memory() can encounter thp only when the thp is being | |
1454 | * worked by memory_failure() and the page lock is not held yet. | |
1455 | * In such case, we yield to memory_failure() and make unpoison fail. | |
1456 | */ | |
e76d30e2 | 1457 | if (!PageHuge(page) && PageTransHuge(page)) { |
0cea3fdc | 1458 | pr_info("MCE: Memory failure is now running on %#lx\n", pfn); |
ead07f6a | 1459 | return 0; |
0cea3fdc WL |
1460 | } |
1461 | ||
f9121153 | 1462 | nr_pages = 1 << compound_order(page); |
c9fbdd5f | 1463 | |
ead07f6a | 1464 | if (!get_hwpoison_page(p)) { |
8c6c2ecb NH |
1465 | /* |
1466 | * Since HWPoisoned hugepage should have non-zero refcount, | |
1467 | * race between memory failure and unpoison seems to happen. | |
1468 | * In such case unpoison fails and memory failure runs | |
1469 | * to the end. | |
1470 | */ | |
1471 | if (PageHuge(page)) { | |
dd73e85f | 1472 | pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn); |
8c6c2ecb NH |
1473 | return 0; |
1474 | } | |
847ce401 | 1475 | if (TestClearPageHWPoison(p)) |
8e30456b | 1476 | num_poisoned_pages_dec(); |
fb46e735 | 1477 | pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn); |
847ce401 WF |
1478 | return 0; |
1479 | } | |
1480 | ||
7eaceacc | 1481 | lock_page(page); |
847ce401 WF |
1482 | /* |
1483 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1484 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1485 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1486 | * the free buddy page pool. | |
1487 | */ | |
c9fbdd5f | 1488 | if (TestClearPageHWPoison(page)) { |
fb46e735 | 1489 | pr_info("MCE: Software-unpoisoned page %#lx\n", pfn); |
8e30456b | 1490 | num_poisoned_pages_sub(nr_pages); |
847ce401 | 1491 | freeit = 1; |
6a90181c NH |
1492 | if (PageHuge(page)) |
1493 | clear_page_hwpoison_huge_page(page); | |
847ce401 WF |
1494 | } |
1495 | unlock_page(page); | |
1496 | ||
665d9da7 | 1497 | put_hwpoison_page(page); |
3ba5eebc | 1498 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) |
665d9da7 | 1499 | put_hwpoison_page(page); |
847ce401 WF |
1500 | |
1501 | return 0; | |
1502 | } | |
1503 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 AK |
1504 | |
1505 | static struct page *new_page(struct page *p, unsigned long private, int **x) | |
1506 | { | |
12686d15 | 1507 | int nid = page_to_nid(p); |
d950b958 NH |
1508 | if (PageHuge(p)) |
1509 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1510 | nid); | |
1511 | else | |
96db800f | 1512 | return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0); |
facb6011 AK |
1513 | } |
1514 | ||
1515 | /* | |
1516 | * Safely get reference count of an arbitrary page. | |
1517 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1518 | * that is not free, and 1 for any other page type. | |
1519 | * For 1 the page is returned with increased page count, otherwise not. | |
1520 | */ | |
af8fae7c | 1521 | static int __get_any_page(struct page *p, unsigned long pfn, int flags) |
facb6011 AK |
1522 | { |
1523 | int ret; | |
1524 | ||
1525 | if (flags & MF_COUNT_INCREASED) | |
1526 | return 1; | |
1527 | ||
d950b958 NH |
1528 | /* |
1529 | * When the target page is a free hugepage, just remove it | |
1530 | * from free hugepage list. | |
1531 | */ | |
ead07f6a | 1532 | if (!get_hwpoison_page(p)) { |
d950b958 | 1533 | if (PageHuge(p)) { |
71dd0b8a | 1534 | pr_info("%s: %#lx free huge page\n", __func__, pfn); |
af8fae7c | 1535 | ret = 0; |
d950b958 | 1536 | } else if (is_free_buddy_page(p)) { |
71dd0b8a | 1537 | pr_info("%s: %#lx free buddy page\n", __func__, pfn); |
facb6011 AK |
1538 | ret = 0; |
1539 | } else { | |
71dd0b8a BP |
1540 | pr_info("%s: %#lx: unknown zero refcount page type %lx\n", |
1541 | __func__, pfn, p->flags); | |
facb6011 AK |
1542 | ret = -EIO; |
1543 | } | |
1544 | } else { | |
1545 | /* Not a free page */ | |
1546 | ret = 1; | |
1547 | } | |
facb6011 AK |
1548 | return ret; |
1549 | } | |
1550 | ||
af8fae7c NH |
1551 | static int get_any_page(struct page *page, unsigned long pfn, int flags) |
1552 | { | |
1553 | int ret = __get_any_page(page, pfn, flags); | |
1554 | ||
1555 | if (ret == 1 && !PageHuge(page) && !PageLRU(page)) { | |
1556 | /* | |
1557 | * Try to free it. | |
1558 | */ | |
665d9da7 | 1559 | put_hwpoison_page(page); |
af8fae7c NH |
1560 | shake_page(page, 1); |
1561 | ||
1562 | /* | |
1563 | * Did it turn free? | |
1564 | */ | |
1565 | ret = __get_any_page(page, pfn, 0); | |
1566 | if (!PageLRU(page)) { | |
4f32be67 | 1567 | /* Drop page reference which is from __get_any_page() */ |
665d9da7 | 1568 | put_hwpoison_page(page); |
af8fae7c NH |
1569 | pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n", |
1570 | pfn, page->flags); | |
1571 | return -EIO; | |
1572 | } | |
1573 | } | |
1574 | return ret; | |
1575 | } | |
1576 | ||
d950b958 NH |
1577 | static int soft_offline_huge_page(struct page *page, int flags) |
1578 | { | |
1579 | int ret; | |
1580 | unsigned long pfn = page_to_pfn(page); | |
1581 | struct page *hpage = compound_head(page); | |
b8ec1cee | 1582 | LIST_HEAD(pagelist); |
d950b958 | 1583 | |
af8fae7c NH |
1584 | /* |
1585 | * This double-check of PageHWPoison is to avoid the race with | |
1586 | * memory_failure(). See also comment in __soft_offline_page(). | |
1587 | */ | |
1588 | lock_page(hpage); | |
0ebff32c | 1589 | if (PageHWPoison(hpage)) { |
af8fae7c | 1590 | unlock_page(hpage); |
665d9da7 | 1591 | put_hwpoison_page(hpage); |
0ebff32c | 1592 | pr_info("soft offline: %#lx hugepage already poisoned\n", pfn); |
af8fae7c | 1593 | return -EBUSY; |
0ebff32c | 1594 | } |
af8fae7c | 1595 | unlock_page(hpage); |
d950b958 | 1596 | |
bcc54222 | 1597 | ret = isolate_huge_page(hpage, &pagelist); |
03613808 WL |
1598 | /* |
1599 | * get_any_page() and isolate_huge_page() takes a refcount each, | |
1600 | * so need to drop one here. | |
1601 | */ | |
665d9da7 | 1602 | put_hwpoison_page(hpage); |
03613808 | 1603 | if (!ret) { |
bcc54222 NH |
1604 | pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn); |
1605 | return -EBUSY; | |
1606 | } | |
1607 | ||
68711a74 | 1608 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, |
b8ec1cee | 1609 | MIGRATE_SYNC, MR_MEMORY_FAILURE); |
d950b958 | 1610 | if (ret) { |
dd73e85f DN |
1611 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
1612 | pfn, ret, page->flags); | |
b8ec1cee NH |
1613 | /* |
1614 | * We know that soft_offline_huge_page() tries to migrate | |
1615 | * only one hugepage pointed to by hpage, so we need not | |
1616 | * run through the pagelist here. | |
1617 | */ | |
1618 | putback_active_hugepage(hpage); | |
1619 | if (ret > 0) | |
1620 | ret = -EIO; | |
af8fae7c | 1621 | } else { |
a49ecbcd JW |
1622 | /* overcommit hugetlb page will be freed to buddy */ |
1623 | if (PageHuge(page)) { | |
1624 | set_page_hwpoison_huge_page(hpage); | |
1625 | dequeue_hwpoisoned_huge_page(hpage); | |
8e30456b | 1626 | num_poisoned_pages_add(1 << compound_order(hpage)); |
a49ecbcd JW |
1627 | } else { |
1628 | SetPageHWPoison(page); | |
8e30456b | 1629 | num_poisoned_pages_inc(); |
a49ecbcd | 1630 | } |
d950b958 | 1631 | } |
d950b958 NH |
1632 | return ret; |
1633 | } | |
1634 | ||
af8fae7c NH |
1635 | static int __soft_offline_page(struct page *page, int flags) |
1636 | { | |
1637 | int ret; | |
1638 | unsigned long pfn = page_to_pfn(page); | |
facb6011 | 1639 | |
facb6011 | 1640 | /* |
af8fae7c NH |
1641 | * Check PageHWPoison again inside page lock because PageHWPoison |
1642 | * is set by memory_failure() outside page lock. Note that | |
1643 | * memory_failure() also double-checks PageHWPoison inside page lock, | |
1644 | * so there's no race between soft_offline_page() and memory_failure(). | |
facb6011 | 1645 | */ |
0ebff32c XQ |
1646 | lock_page(page); |
1647 | wait_on_page_writeback(page); | |
af8fae7c NH |
1648 | if (PageHWPoison(page)) { |
1649 | unlock_page(page); | |
665d9da7 | 1650 | put_hwpoison_page(page); |
af8fae7c NH |
1651 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
1652 | return -EBUSY; | |
1653 | } | |
facb6011 AK |
1654 | /* |
1655 | * Try to invalidate first. This should work for | |
1656 | * non dirty unmapped page cache pages. | |
1657 | */ | |
1658 | ret = invalidate_inode_page(page); | |
1659 | unlock_page(page); | |
facb6011 | 1660 | /* |
facb6011 AK |
1661 | * RED-PEN would be better to keep it isolated here, but we |
1662 | * would need to fix isolation locking first. | |
1663 | */ | |
facb6011 | 1664 | if (ret == 1) { |
665d9da7 | 1665 | put_hwpoison_page(page); |
fb46e735 | 1666 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
af8fae7c | 1667 | SetPageHWPoison(page); |
8e30456b | 1668 | num_poisoned_pages_inc(); |
af8fae7c | 1669 | return 0; |
facb6011 AK |
1670 | } |
1671 | ||
1672 | /* | |
1673 | * Simple invalidation didn't work. | |
1674 | * Try to migrate to a new page instead. migrate.c | |
1675 | * handles a large number of cases for us. | |
1676 | */ | |
1677 | ret = isolate_lru_page(page); | |
bd486285 KK |
1678 | /* |
1679 | * Drop page reference which is came from get_any_page() | |
1680 | * successful isolate_lru_page() already took another one. | |
1681 | */ | |
665d9da7 | 1682 | put_hwpoison_page(page); |
facb6011 AK |
1683 | if (!ret) { |
1684 | LIST_HEAD(pagelist); | |
5db8a73a | 1685 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
9c620e2b | 1686 | page_is_file_cache(page)); |
facb6011 | 1687 | list_add(&page->lru, &pagelist); |
68711a74 | 1688 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, |
9c620e2b | 1689 | MIGRATE_SYNC, MR_MEMORY_FAILURE); |
facb6011 | 1690 | if (ret) { |
59c82b70 JK |
1691 | if (!list_empty(&pagelist)) { |
1692 | list_del(&page->lru); | |
1693 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
1694 | page_is_file_cache(page)); | |
1695 | putback_lru_page(page); | |
1696 | } | |
1697 | ||
fb46e735 | 1698 | pr_info("soft offline: %#lx: migration failed %d, type %lx\n", |
facb6011 AK |
1699 | pfn, ret, page->flags); |
1700 | if (ret > 0) | |
1701 | ret = -EIO; | |
1702 | } | |
1703 | } else { | |
fb46e735 | 1704 | pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n", |
dd73e85f | 1705 | pfn, ret, page_count(page), page->flags); |
facb6011 | 1706 | } |
facb6011 AK |
1707 | return ret; |
1708 | } | |
86e05773 WL |
1709 | |
1710 | /** | |
1711 | * soft_offline_page - Soft offline a page. | |
1712 | * @page: page to offline | |
1713 | * @flags: flags. Same as memory_failure(). | |
1714 | * | |
1715 | * Returns 0 on success, otherwise negated errno. | |
1716 | * | |
1717 | * Soft offline a page, by migration or invalidation, | |
1718 | * without killing anything. This is for the case when | |
1719 | * a page is not corrupted yet (so it's still valid to access), | |
1720 | * but has had a number of corrected errors and is better taken | |
1721 | * out. | |
1722 | * | |
1723 | * The actual policy on when to do that is maintained by | |
1724 | * user space. | |
1725 | * | |
1726 | * This should never impact any application or cause data loss, | |
1727 | * however it might take some time. | |
1728 | * | |
1729 | * This is not a 100% solution for all memory, but tries to be | |
1730 | * ``good enough'' for the majority of memory. | |
1731 | */ | |
1732 | int soft_offline_page(struct page *page, int flags) | |
1733 | { | |
1734 | int ret; | |
1735 | unsigned long pfn = page_to_pfn(page); | |
668f9abb | 1736 | struct page *hpage = compound_head(page); |
86e05773 WL |
1737 | |
1738 | if (PageHWPoison(page)) { | |
1739 | pr_info("soft offline: %#lx page already poisoned\n", pfn); | |
1e0e635b | 1740 | if (flags & MF_COUNT_INCREASED) |
665d9da7 | 1741 | put_hwpoison_page(page); |
86e05773 WL |
1742 | return -EBUSY; |
1743 | } | |
1744 | if (!PageHuge(page) && PageTransHuge(hpage)) { | |
1745 | if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) { | |
1746 | pr_info("soft offline: %#lx: failed to split THP\n", | |
1747 | pfn); | |
7d1900c7 | 1748 | if (flags & MF_COUNT_INCREASED) |
665d9da7 | 1749 | put_hwpoison_page(page); |
86e05773 WL |
1750 | return -EBUSY; |
1751 | } | |
1752 | } | |
1753 | ||
bfc8c901 | 1754 | get_online_mems(); |
03b61ff3 | 1755 | |
86e05773 | 1756 | ret = get_any_page(page, pfn, flags); |
bfc8c901 | 1757 | put_online_mems(); |
03b61ff3 | 1758 | if (ret > 0) { /* for in-use pages */ |
86e05773 WL |
1759 | if (PageHuge(page)) |
1760 | ret = soft_offline_huge_page(page, flags); | |
1761 | else | |
1762 | ret = __soft_offline_page(page, flags); | |
03b61ff3 | 1763 | } else if (ret == 0) { /* for free pages */ |
86e05773 WL |
1764 | if (PageHuge(page)) { |
1765 | set_page_hwpoison_huge_page(hpage); | |
602498f9 | 1766 | if (!dequeue_hwpoisoned_huge_page(hpage)) |
8e30456b | 1767 | num_poisoned_pages_add(1 << compound_order(hpage)); |
86e05773 | 1768 | } else { |
602498f9 | 1769 | if (!TestSetPageHWPoison(page)) |
8e30456b | 1770 | num_poisoned_pages_inc(); |
86e05773 WL |
1771 | } |
1772 | } | |
86e05773 WL |
1773 | return ret; |
1774 | } |