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> |
3f07c014 | 43 | #include <linux/sched/signal.h> |
29930025 | 44 | #include <linux/sched/task.h> |
01e00f88 | 45 | #include <linux/ksm.h> |
6a46079c | 46 | #include <linux/rmap.h> |
b9e15baf | 47 | #include <linux/export.h> |
6a46079c AK |
48 | #include <linux/pagemap.h> |
49 | #include <linux/swap.h> | |
50 | #include <linux/backing-dev.h> | |
facb6011 | 51 | #include <linux/migrate.h> |
facb6011 | 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> |
6100e34b | 58 | #include <linux/memremap.h> |
ea8f5fb8 | 59 | #include <linux/kfifo.h> |
a5f65109 | 60 | #include <linux/ratelimit.h> |
d4ae9916 | 61 | #include <linux/page-isolation.h> |
6a46079c | 62 | #include "internal.h" |
97f0b134 | 63 | #include "ras/ras_event.h" |
6a46079c AK |
64 | |
65 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
66 | ||
67 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
68 | ||
293c07e3 | 69 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 70 | |
27df5068 AK |
71 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
72 | ||
1bfe5feb | 73 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
74 | u32 hwpoison_filter_dev_major = ~0U; |
75 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
76 | u64 hwpoison_filter_flags_mask; |
77 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 78 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
79 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
80 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
81 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
82 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
83 | |
84 | static int hwpoison_filter_dev(struct page *p) | |
85 | { | |
86 | struct address_space *mapping; | |
87 | dev_t dev; | |
88 | ||
89 | if (hwpoison_filter_dev_major == ~0U && | |
90 | hwpoison_filter_dev_minor == ~0U) | |
91 | return 0; | |
92 | ||
93 | /* | |
1c80b990 | 94 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
95 | */ |
96 | if (PageSlab(p)) | |
97 | return -EINVAL; | |
98 | ||
99 | mapping = page_mapping(p); | |
100 | if (mapping == NULL || mapping->host == NULL) | |
101 | return -EINVAL; | |
102 | ||
103 | dev = mapping->host->i_sb->s_dev; | |
104 | if (hwpoison_filter_dev_major != ~0U && | |
105 | hwpoison_filter_dev_major != MAJOR(dev)) | |
106 | return -EINVAL; | |
107 | if (hwpoison_filter_dev_minor != ~0U && | |
108 | hwpoison_filter_dev_minor != MINOR(dev)) | |
109 | return -EINVAL; | |
110 | ||
111 | return 0; | |
112 | } | |
113 | ||
478c5ffc WF |
114 | static int hwpoison_filter_flags(struct page *p) |
115 | { | |
116 | if (!hwpoison_filter_flags_mask) | |
117 | return 0; | |
118 | ||
119 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
120 | hwpoison_filter_flags_value) | |
121 | return 0; | |
122 | else | |
123 | return -EINVAL; | |
124 | } | |
125 | ||
4fd466eb AK |
126 | /* |
127 | * This allows stress tests to limit test scope to a collection of tasks | |
128 | * by putting them under some memcg. This prevents killing unrelated/important | |
129 | * processes such as /sbin/init. Note that the target task may share clean | |
130 | * pages with init (eg. libc text), which is harmless. If the target task | |
131 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
132 | * is also included in the memcg. At last, due to race conditions this filter | |
133 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
134 | * a freed page. | |
135 | */ | |
94a59fb3 | 136 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
137 | u64 hwpoison_filter_memcg; |
138 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
139 | static int hwpoison_filter_task(struct page *p) | |
140 | { | |
4fd466eb AK |
141 | if (!hwpoison_filter_memcg) |
142 | return 0; | |
143 | ||
94a59fb3 | 144 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
145 | return -EINVAL; |
146 | ||
147 | return 0; | |
148 | } | |
149 | #else | |
150 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
151 | #endif | |
152 | ||
7c116f2b WF |
153 | int hwpoison_filter(struct page *p) |
154 | { | |
1bfe5feb HL |
155 | if (!hwpoison_filter_enable) |
156 | return 0; | |
157 | ||
7c116f2b WF |
158 | if (hwpoison_filter_dev(p)) |
159 | return -EINVAL; | |
160 | ||
478c5ffc WF |
161 | if (hwpoison_filter_flags(p)) |
162 | return -EINVAL; | |
163 | ||
4fd466eb AK |
164 | if (hwpoison_filter_task(p)) |
165 | return -EINVAL; | |
166 | ||
7c116f2b WF |
167 | return 0; |
168 | } | |
27df5068 AK |
169 | #else |
170 | int hwpoison_filter(struct page *p) | |
171 | { | |
172 | return 0; | |
173 | } | |
174 | #endif | |
175 | ||
7c116f2b WF |
176 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
177 | ||
ae1139ec DW |
178 | /* |
179 | * Kill all processes that have a poisoned page mapped and then isolate | |
180 | * the page. | |
181 | * | |
182 | * General strategy: | |
183 | * Find all processes having the page mapped and kill them. | |
184 | * But we keep a page reference around so that the page is not | |
185 | * actually freed yet. | |
186 | * Then stash the page away | |
187 | * | |
188 | * There's no convenient way to get back to mapped processes | |
189 | * from the VMAs. So do a brute-force search over all | |
190 | * running processes. | |
191 | * | |
192 | * Remember that machine checks are not common (or rather | |
193 | * if they are common you have other problems), so this shouldn't | |
194 | * be a performance issue. | |
195 | * | |
196 | * Also there are some races possible while we get from the | |
197 | * error detection to actually handle it. | |
198 | */ | |
199 | ||
200 | struct to_kill { | |
201 | struct list_head nd; | |
202 | struct task_struct *tsk; | |
203 | unsigned long addr; | |
204 | short size_shift; | |
205 | char addr_valid; | |
206 | }; | |
207 | ||
6a46079c | 208 | /* |
7329bbeb TL |
209 | * Send all the processes who have the page mapped a signal. |
210 | * ``action optional'' if they are not immediately affected by the error | |
211 | * ``action required'' if error happened in current execution context | |
6a46079c | 212 | */ |
ae1139ec | 213 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 214 | { |
ae1139ec DW |
215 | struct task_struct *t = tk->tsk; |
216 | short addr_lsb = tk->size_shift; | |
6a46079c AK |
217 | int ret; |
218 | ||
495367c0 CY |
219 | pr_err("Memory failure: %#lx: Killing %s:%d due to hardware memory corruption\n", |
220 | pfn, t->comm, t->pid); | |
7329bbeb | 221 | |
a70ffcac | 222 | if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) { |
ae1139ec | 223 | ret = force_sig_mceerr(BUS_MCEERR_AR, (void __user *)tk->addr, |
c0f45555 | 224 | addr_lsb, current); |
7329bbeb TL |
225 | } else { |
226 | /* | |
227 | * Don't use force here, it's convenient if the signal | |
228 | * can be temporarily blocked. | |
229 | * This could cause a loop when the user sets SIGBUS | |
230 | * to SIG_IGN, but hopefully no one will do that? | |
231 | */ | |
ae1139ec | 232 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
c0f45555 | 233 | addr_lsb, t); /* synchronous? */ |
7329bbeb | 234 | } |
6a46079c | 235 | if (ret < 0) |
495367c0 | 236 | pr_info("Memory failure: Error sending signal to %s:%d: %d\n", |
1170532b | 237 | t->comm, t->pid, ret); |
6a46079c AK |
238 | return ret; |
239 | } | |
240 | ||
588f9ce6 AK |
241 | /* |
242 | * When a unknown page type is encountered drain as many buffers as possible | |
243 | * in the hope to turn the page into a LRU or free page, which we can handle. | |
244 | */ | |
facb6011 | 245 | void shake_page(struct page *p, int access) |
588f9ce6 | 246 | { |
8bcb74de NH |
247 | if (PageHuge(p)) |
248 | return; | |
249 | ||
588f9ce6 AK |
250 | if (!PageSlab(p)) { |
251 | lru_add_drain_all(); | |
252 | if (PageLRU(p)) | |
253 | return; | |
c0554329 | 254 | drain_all_pages(page_zone(p)); |
588f9ce6 AK |
255 | if (PageLRU(p) || is_free_buddy_page(p)) |
256 | return; | |
257 | } | |
facb6011 | 258 | |
588f9ce6 | 259 | /* |
6b4f7799 JW |
260 | * Only call shrink_node_slabs here (which would also shrink |
261 | * other caches) if access is not potentially fatal. | |
588f9ce6 | 262 | */ |
cb731d6c VD |
263 | if (access) |
264 | drop_slab_node(page_to_nid(p)); | |
588f9ce6 AK |
265 | } |
266 | EXPORT_SYMBOL_GPL(shake_page); | |
267 | ||
6100e34b DW |
268 | static unsigned long dev_pagemap_mapping_shift(struct page *page, |
269 | struct vm_area_struct *vma) | |
270 | { | |
271 | unsigned long address = vma_address(page, vma); | |
272 | pgd_t *pgd; | |
273 | p4d_t *p4d; | |
274 | pud_t *pud; | |
275 | pmd_t *pmd; | |
276 | pte_t *pte; | |
277 | ||
278 | pgd = pgd_offset(vma->vm_mm, address); | |
279 | if (!pgd_present(*pgd)) | |
280 | return 0; | |
281 | p4d = p4d_offset(pgd, address); | |
282 | if (!p4d_present(*p4d)) | |
283 | return 0; | |
284 | pud = pud_offset(p4d, address); | |
285 | if (!pud_present(*pud)) | |
286 | return 0; | |
287 | if (pud_devmap(*pud)) | |
288 | return PUD_SHIFT; | |
289 | pmd = pmd_offset(pud, address); | |
290 | if (!pmd_present(*pmd)) | |
291 | return 0; | |
292 | if (pmd_devmap(*pmd)) | |
293 | return PMD_SHIFT; | |
294 | pte = pte_offset_map(pmd, address); | |
295 | if (!pte_present(*pte)) | |
296 | return 0; | |
297 | if (pte_devmap(*pte)) | |
298 | return PAGE_SHIFT; | |
299 | return 0; | |
300 | } | |
6a46079c AK |
301 | |
302 | /* | |
303 | * Failure handling: if we can't find or can't kill a process there's | |
304 | * not much we can do. We just print a message and ignore otherwise. | |
305 | */ | |
306 | ||
307 | /* | |
308 | * Schedule a process for later kill. | |
309 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
310 | * TBD would GFP_NOIO be enough? | |
311 | */ | |
312 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
313 | struct vm_area_struct *vma, | |
314 | struct list_head *to_kill, | |
315 | struct to_kill **tkc) | |
316 | { | |
317 | struct to_kill *tk; | |
318 | ||
319 | if (*tkc) { | |
320 | tk = *tkc; | |
321 | *tkc = NULL; | |
322 | } else { | |
323 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); | |
324 | if (!tk) { | |
495367c0 | 325 | pr_err("Memory failure: Out of memory while machine check handling\n"); |
6a46079c AK |
326 | return; |
327 | } | |
328 | } | |
329 | tk->addr = page_address_in_vma(p, vma); | |
330 | tk->addr_valid = 1; | |
6100e34b DW |
331 | if (is_zone_device_page(p)) |
332 | tk->size_shift = dev_pagemap_mapping_shift(p, vma); | |
333 | else | |
334 | tk->size_shift = compound_order(compound_head(p)) + PAGE_SHIFT; | |
6a46079c AK |
335 | |
336 | /* | |
337 | * In theory we don't have to kill when the page was | |
338 | * munmaped. But it could be also a mremap. Since that's | |
339 | * likely very rare kill anyways just out of paranoia, but use | |
340 | * a SIGKILL because the error is not contained anymore. | |
341 | */ | |
6100e34b | 342 | if (tk->addr == -EFAULT || tk->size_shift == 0) { |
495367c0 | 343 | pr_info("Memory failure: Unable to find user space address %lx in %s\n", |
6a46079c AK |
344 | page_to_pfn(p), tsk->comm); |
345 | tk->addr_valid = 0; | |
346 | } | |
347 | get_task_struct(tsk); | |
348 | tk->tsk = tsk; | |
349 | list_add_tail(&tk->nd, to_kill); | |
350 | } | |
351 | ||
352 | /* | |
353 | * Kill the processes that have been collected earlier. | |
354 | * | |
355 | * Only do anything when DOIT is set, otherwise just free the list | |
356 | * (this is used for clean pages which do not need killing) | |
357 | * Also when FAIL is set do a force kill because something went | |
358 | * wrong earlier. | |
359 | */ | |
ae1139ec DW |
360 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
361 | unsigned long pfn, int flags) | |
6a46079c AK |
362 | { |
363 | struct to_kill *tk, *next; | |
364 | ||
365 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 366 | if (forcekill) { |
6a46079c | 367 | /* |
af901ca1 | 368 | * In case something went wrong with munmapping |
6a46079c AK |
369 | * make sure the process doesn't catch the |
370 | * signal and then access the memory. Just kill it. | |
6a46079c AK |
371 | */ |
372 | if (fail || tk->addr_valid == 0) { | |
495367c0 | 373 | pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 374 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c AK |
375 | force_sig(SIGKILL, tk->tsk); |
376 | } | |
377 | ||
378 | /* | |
379 | * In theory the process could have mapped | |
380 | * something else on the address in-between. We could | |
381 | * check for that, but we need to tell the | |
382 | * process anyways. | |
383 | */ | |
ae1139ec | 384 | else if (kill_proc(tk, pfn, flags) < 0) |
495367c0 | 385 | pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 386 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c AK |
387 | } |
388 | put_task_struct(tk->tsk); | |
389 | kfree(tk); | |
390 | } | |
391 | } | |
392 | ||
3ba08129 NH |
393 | /* |
394 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
395 | * on behalf of the thread group. Return task_struct of the (first found) | |
396 | * dedicated thread if found, and return NULL otherwise. | |
397 | * | |
398 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
399 | * have to call rcu_read_lock/unlock() in this function. | |
400 | */ | |
401 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 402 | { |
3ba08129 NH |
403 | struct task_struct *t; |
404 | ||
405 | for_each_thread(tsk, t) | |
406 | if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY)) | |
407 | return t; | |
408 | return NULL; | |
409 | } | |
410 | ||
411 | /* | |
412 | * Determine whether a given process is "early kill" process which expects | |
413 | * to be signaled when some page under the process is hwpoisoned. | |
414 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
415 | * specified) if the process is "early kill," and otherwise returns NULL. | |
416 | */ | |
417 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
418 | int force_early) | |
419 | { | |
420 | struct task_struct *t; | |
6a46079c | 421 | if (!tsk->mm) |
3ba08129 | 422 | return NULL; |
74614de1 | 423 | if (force_early) |
3ba08129 NH |
424 | return tsk; |
425 | t = find_early_kill_thread(tsk); | |
426 | if (t) | |
427 | return t; | |
428 | if (sysctl_memory_failure_early_kill) | |
429 | return tsk; | |
430 | return NULL; | |
6a46079c AK |
431 | } |
432 | ||
433 | /* | |
434 | * Collect processes when the error hit an anonymous page. | |
435 | */ | |
436 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
74614de1 | 437 | struct to_kill **tkc, int force_early) |
6a46079c AK |
438 | { |
439 | struct vm_area_struct *vma; | |
440 | struct task_struct *tsk; | |
441 | struct anon_vma *av; | |
bf181b9f | 442 | pgoff_t pgoff; |
6a46079c | 443 | |
4fc3f1d6 | 444 | av = page_lock_anon_vma_read(page); |
6a46079c | 445 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
446 | return; |
447 | ||
a0f7a756 | 448 | pgoff = page_to_pgoff(page); |
9b679320 | 449 | read_lock(&tasklist_lock); |
6a46079c | 450 | for_each_process (tsk) { |
5beb4930 | 451 | struct anon_vma_chain *vmac; |
3ba08129 | 452 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 453 | |
3ba08129 | 454 | if (!t) |
6a46079c | 455 | continue; |
bf181b9f ML |
456 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
457 | pgoff, pgoff) { | |
5beb4930 | 458 | vma = vmac->vma; |
6a46079c AK |
459 | if (!page_mapped_in_vma(page, vma)) |
460 | continue; | |
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); |
4fc3f1d6 | 466 | page_unlock_anon_vma_read(av); |
6a46079c AK |
467 | } |
468 | ||
469 | /* | |
470 | * Collect processes when the error hit a file mapped page. | |
471 | */ | |
472 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
74614de1 | 473 | struct to_kill **tkc, int force_early) |
6a46079c AK |
474 | { |
475 | struct vm_area_struct *vma; | |
476 | struct task_struct *tsk; | |
6a46079c AK |
477 | struct address_space *mapping = page->mapping; |
478 | ||
d28eb9c8 | 479 | i_mmap_lock_read(mapping); |
9b679320 | 480 | read_lock(&tasklist_lock); |
6a46079c | 481 | for_each_process(tsk) { |
a0f7a756 | 482 | pgoff_t pgoff = page_to_pgoff(page); |
3ba08129 | 483 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 484 | |
3ba08129 | 485 | if (!t) |
6a46079c | 486 | continue; |
6b2dbba8 | 487 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
488 | pgoff) { |
489 | /* | |
490 | * Send early kill signal to tasks where a vma covers | |
491 | * the page but the corrupted page is not necessarily | |
492 | * mapped it in its pte. | |
493 | * Assume applications who requested early kill want | |
494 | * to be informed of all such data corruptions. | |
495 | */ | |
3ba08129 NH |
496 | if (vma->vm_mm == t->mm) |
497 | add_to_kill(t, page, vma, to_kill, tkc); | |
6a46079c AK |
498 | } |
499 | } | |
6a46079c | 500 | read_unlock(&tasklist_lock); |
d28eb9c8 | 501 | i_mmap_unlock_read(mapping); |
6a46079c AK |
502 | } |
503 | ||
504 | /* | |
505 | * Collect the processes who have the corrupted page mapped to kill. | |
506 | * This is done in two steps for locking reasons. | |
507 | * First preallocate one tokill structure outside the spin locks, | |
508 | * so that we can kill at least one process reasonably reliable. | |
509 | */ | |
74614de1 TL |
510 | static void collect_procs(struct page *page, struct list_head *tokill, |
511 | int force_early) | |
6a46079c AK |
512 | { |
513 | struct to_kill *tk; | |
514 | ||
515 | if (!page->mapping) | |
516 | return; | |
517 | ||
518 | tk = kmalloc(sizeof(struct to_kill), GFP_NOIO); | |
519 | if (!tk) | |
520 | return; | |
521 | if (PageAnon(page)) | |
74614de1 | 522 | collect_procs_anon(page, tokill, &tk, force_early); |
6a46079c | 523 | else |
74614de1 | 524 | collect_procs_file(page, tokill, &tk, force_early); |
6a46079c AK |
525 | kfree(tk); |
526 | } | |
527 | ||
6a46079c | 528 | static const char *action_name[] = { |
cc637b17 XX |
529 | [MF_IGNORED] = "Ignored", |
530 | [MF_FAILED] = "Failed", | |
531 | [MF_DELAYED] = "Delayed", | |
532 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
533 | }; |
534 | ||
535 | static const char * const action_page_types[] = { | |
cc637b17 XX |
536 | [MF_MSG_KERNEL] = "reserved kernel page", |
537 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
538 | [MF_MSG_SLAB] = "kernel slab page", | |
539 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
540 | [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", | |
541 | [MF_MSG_HUGE] = "huge page", | |
542 | [MF_MSG_FREE_HUGE] = "free huge page", | |
31286a84 | 543 | [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page", |
cc637b17 XX |
544 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", |
545 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
546 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
547 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
548 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
549 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
550 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
551 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
552 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
553 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
554 | [MF_MSG_BUDDY] = "free buddy page", | |
555 | [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", | |
6100e34b | 556 | [MF_MSG_DAX] = "dax page", |
cc637b17 | 557 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
558 | }; |
559 | ||
dc2a1cbf WF |
560 | /* |
561 | * XXX: It is possible that a page is isolated from LRU cache, | |
562 | * and then kept in swap cache or failed to remove from page cache. | |
563 | * The page count will stop it from being freed by unpoison. | |
564 | * Stress tests should be aware of this memory leak problem. | |
565 | */ | |
566 | static int delete_from_lru_cache(struct page *p) | |
567 | { | |
568 | if (!isolate_lru_page(p)) { | |
569 | /* | |
570 | * Clear sensible page flags, so that the buddy system won't | |
571 | * complain when the page is unpoison-and-freed. | |
572 | */ | |
573 | ClearPageActive(p); | |
574 | ClearPageUnevictable(p); | |
18365225 MH |
575 | |
576 | /* | |
577 | * Poisoned page might never drop its ref count to 0 so we have | |
578 | * to uncharge it manually from its memcg. | |
579 | */ | |
580 | mem_cgroup_uncharge(p); | |
581 | ||
dc2a1cbf WF |
582 | /* |
583 | * drop the page count elevated by isolate_lru_page() | |
584 | */ | |
09cbfeaf | 585 | put_page(p); |
dc2a1cbf WF |
586 | return 0; |
587 | } | |
588 | return -EIO; | |
589 | } | |
590 | ||
78bb9203 NH |
591 | static int truncate_error_page(struct page *p, unsigned long pfn, |
592 | struct address_space *mapping) | |
593 | { | |
594 | int ret = MF_FAILED; | |
595 | ||
596 | if (mapping->a_ops->error_remove_page) { | |
597 | int err = mapping->a_ops->error_remove_page(mapping, p); | |
598 | ||
599 | if (err != 0) { | |
600 | pr_info("Memory failure: %#lx: Failed to punch page: %d\n", | |
601 | pfn, err); | |
602 | } else if (page_has_private(p) && | |
603 | !try_to_release_page(p, GFP_NOIO)) { | |
604 | pr_info("Memory failure: %#lx: failed to release buffers\n", | |
605 | pfn); | |
606 | } else { | |
607 | ret = MF_RECOVERED; | |
608 | } | |
609 | } else { | |
610 | /* | |
611 | * If the file system doesn't support it just invalidate | |
612 | * This fails on dirty or anything with private pages | |
613 | */ | |
614 | if (invalidate_inode_page(p)) | |
615 | ret = MF_RECOVERED; | |
616 | else | |
617 | pr_info("Memory failure: %#lx: Failed to invalidate\n", | |
618 | pfn); | |
619 | } | |
620 | ||
621 | return ret; | |
622 | } | |
623 | ||
6a46079c AK |
624 | /* |
625 | * Error hit kernel page. | |
626 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
627 | * could be more sophisticated. | |
628 | */ | |
629 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c | 630 | { |
cc637b17 | 631 | return MF_IGNORED; |
6a46079c AK |
632 | } |
633 | ||
634 | /* | |
635 | * Page in unknown state. Do nothing. | |
636 | */ | |
637 | static int me_unknown(struct page *p, unsigned long pfn) | |
638 | { | |
495367c0 | 639 | pr_err("Memory failure: %#lx: Unknown page state\n", pfn); |
cc637b17 | 640 | return MF_FAILED; |
6a46079c AK |
641 | } |
642 | ||
6a46079c AK |
643 | /* |
644 | * Clean (or cleaned) page cache page. | |
645 | */ | |
646 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
647 | { | |
6a46079c AK |
648 | struct address_space *mapping; |
649 | ||
dc2a1cbf WF |
650 | delete_from_lru_cache(p); |
651 | ||
6a46079c AK |
652 | /* |
653 | * For anonymous pages we're done the only reference left | |
654 | * should be the one m_f() holds. | |
655 | */ | |
656 | if (PageAnon(p)) | |
cc637b17 | 657 | return MF_RECOVERED; |
6a46079c AK |
658 | |
659 | /* | |
660 | * Now truncate the page in the page cache. This is really | |
661 | * more like a "temporary hole punch" | |
662 | * Don't do this for block devices when someone else | |
663 | * has a reference, because it could be file system metadata | |
664 | * and that's not safe to truncate. | |
665 | */ | |
666 | mapping = page_mapping(p); | |
667 | if (!mapping) { | |
668 | /* | |
669 | * Page has been teared down in the meanwhile | |
670 | */ | |
cc637b17 | 671 | return MF_FAILED; |
6a46079c AK |
672 | } |
673 | ||
674 | /* | |
675 | * Truncation is a bit tricky. Enable it per file system for now. | |
676 | * | |
677 | * Open: to take i_mutex or not for this? Right now we don't. | |
678 | */ | |
78bb9203 | 679 | return truncate_error_page(p, pfn, mapping); |
6a46079c AK |
680 | } |
681 | ||
682 | /* | |
549543df | 683 | * Dirty pagecache page |
6a46079c AK |
684 | * Issues: when the error hit a hole page the error is not properly |
685 | * propagated. | |
686 | */ | |
687 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
688 | { | |
689 | struct address_space *mapping = page_mapping(p); | |
690 | ||
691 | SetPageError(p); | |
692 | /* TBD: print more information about the file. */ | |
693 | if (mapping) { | |
694 | /* | |
695 | * IO error will be reported by write(), fsync(), etc. | |
696 | * who check the mapping. | |
697 | * This way the application knows that something went | |
698 | * wrong with its dirty file data. | |
699 | * | |
700 | * There's one open issue: | |
701 | * | |
702 | * The EIO will be only reported on the next IO | |
703 | * operation and then cleared through the IO map. | |
704 | * Normally Linux has two mechanisms to pass IO error | |
705 | * first through the AS_EIO flag in the address space | |
706 | * and then through the PageError flag in the page. | |
707 | * Since we drop pages on memory failure handling the | |
708 | * only mechanism open to use is through AS_AIO. | |
709 | * | |
710 | * This has the disadvantage that it gets cleared on | |
711 | * the first operation that returns an error, while | |
712 | * the PageError bit is more sticky and only cleared | |
713 | * when the page is reread or dropped. If an | |
714 | * application assumes it will always get error on | |
715 | * fsync, but does other operations on the fd before | |
25985edc | 716 | * and the page is dropped between then the error |
6a46079c AK |
717 | * will not be properly reported. |
718 | * | |
719 | * This can already happen even without hwpoisoned | |
720 | * pages: first on metadata IO errors (which only | |
721 | * report through AS_EIO) or when the page is dropped | |
722 | * at the wrong time. | |
723 | * | |
724 | * So right now we assume that the application DTRT on | |
725 | * the first EIO, but we're not worse than other parts | |
726 | * of the kernel. | |
727 | */ | |
af21bfaf | 728 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
729 | } |
730 | ||
731 | return me_pagecache_clean(p, pfn); | |
732 | } | |
733 | ||
734 | /* | |
735 | * Clean and dirty swap cache. | |
736 | * | |
737 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
738 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
739 | * referenced concurrently by 2 types of PTEs: | |
740 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
741 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
742 | * and then | |
743 | * - clear dirty bit to prevent IO | |
744 | * - remove from LRU | |
745 | * - but keep in the swap cache, so that when we return to it on | |
746 | * a later page fault, we know the application is accessing | |
747 | * corrupted data and shall be killed (we installed simple | |
748 | * interception code in do_swap_page to catch it). | |
749 | * | |
750 | * Clean swap cache pages can be directly isolated. A later page fault will | |
751 | * bring in the known good data from disk. | |
752 | */ | |
753 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
754 | { | |
6a46079c AK |
755 | ClearPageDirty(p); |
756 | /* Trigger EIO in shmem: */ | |
757 | ClearPageUptodate(p); | |
758 | ||
dc2a1cbf | 759 | if (!delete_from_lru_cache(p)) |
cc637b17 | 760 | return MF_DELAYED; |
dc2a1cbf | 761 | else |
cc637b17 | 762 | return MF_FAILED; |
6a46079c AK |
763 | } |
764 | ||
765 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
766 | { | |
6a46079c | 767 | delete_from_swap_cache(p); |
e43c3afb | 768 | |
dc2a1cbf | 769 | if (!delete_from_lru_cache(p)) |
cc637b17 | 770 | return MF_RECOVERED; |
dc2a1cbf | 771 | else |
cc637b17 | 772 | return MF_FAILED; |
6a46079c AK |
773 | } |
774 | ||
775 | /* | |
776 | * Huge pages. Needs work. | |
777 | * Issues: | |
93f70f90 NH |
778 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
779 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
780 | */ |
781 | static int me_huge_page(struct page *p, unsigned long pfn) | |
782 | { | |
6de2b1aa | 783 | int res = 0; |
93f70f90 | 784 | struct page *hpage = compound_head(p); |
78bb9203 | 785 | struct address_space *mapping; |
2491ffee NH |
786 | |
787 | if (!PageHuge(hpage)) | |
788 | return MF_DELAYED; | |
789 | ||
78bb9203 NH |
790 | mapping = page_mapping(hpage); |
791 | if (mapping) { | |
792 | res = truncate_error_page(hpage, pfn, mapping); | |
793 | } else { | |
794 | unlock_page(hpage); | |
795 | /* | |
796 | * migration entry prevents later access on error anonymous | |
797 | * hugepage, so we can free and dissolve it into buddy to | |
798 | * save healthy subpages. | |
799 | */ | |
800 | if (PageAnon(hpage)) | |
801 | put_page(hpage); | |
802 | dissolve_free_huge_page(p); | |
803 | res = MF_RECOVERED; | |
804 | lock_page(hpage); | |
93f70f90 | 805 | } |
78bb9203 NH |
806 | |
807 | return res; | |
6a46079c AK |
808 | } |
809 | ||
810 | /* | |
811 | * Various page states we can handle. | |
812 | * | |
813 | * A page state is defined by its current page->flags bits. | |
814 | * The table matches them in order and calls the right handler. | |
815 | * | |
816 | * This is quite tricky because we can access page at any time | |
25985edc | 817 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
818 | * |
819 | * This is not complete. More states could be added. | |
820 | * For any missing state don't attempt recovery. | |
821 | */ | |
822 | ||
823 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 824 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
825 | #define unevict (1UL << PG_unevictable) |
826 | #define mlock (1UL << PG_mlocked) | |
827 | #define writeback (1UL << PG_writeback) | |
828 | #define lru (1UL << PG_lru) | |
6a46079c | 829 | #define head (1UL << PG_head) |
6a46079c | 830 | #define slab (1UL << PG_slab) |
6a46079c AK |
831 | #define reserved (1UL << PG_reserved) |
832 | ||
833 | static struct page_state { | |
834 | unsigned long mask; | |
835 | unsigned long res; | |
cc637b17 | 836 | enum mf_action_page_type type; |
6a46079c AK |
837 | int (*action)(struct page *p, unsigned long pfn); |
838 | } error_states[] = { | |
cc637b17 | 839 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
840 | /* |
841 | * free pages are specially detected outside this table: | |
842 | * PG_buddy pages only make a small fraction of all free pages. | |
843 | */ | |
6a46079c AK |
844 | |
845 | /* | |
846 | * Could in theory check if slab page is free or if we can drop | |
847 | * currently unused objects without touching them. But just | |
848 | * treat it as standard kernel for now. | |
849 | */ | |
cc637b17 | 850 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 851 | |
cc637b17 | 852 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 853 | |
cc637b17 XX |
854 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
855 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 856 | |
cc637b17 XX |
857 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
858 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 859 | |
cc637b17 XX |
860 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
861 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 862 | |
cc637b17 XX |
863 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
864 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
865 | |
866 | /* | |
867 | * Catchall entry: must be at end. | |
868 | */ | |
cc637b17 | 869 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
870 | }; |
871 | ||
2326c467 AK |
872 | #undef dirty |
873 | #undef sc | |
874 | #undef unevict | |
875 | #undef mlock | |
876 | #undef writeback | |
877 | #undef lru | |
2326c467 | 878 | #undef head |
2326c467 AK |
879 | #undef slab |
880 | #undef reserved | |
881 | ||
ff604cf6 NH |
882 | /* |
883 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
884 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
885 | */ | |
cc3e2af4 XX |
886 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
887 | enum mf_result result) | |
6a46079c | 888 | { |
97f0b134 XX |
889 | trace_memory_failure_event(pfn, type, result); |
890 | ||
495367c0 | 891 | pr_err("Memory failure: %#lx: recovery action for %s: %s\n", |
64d37a2b | 892 | pfn, action_page_types[type], action_name[result]); |
6a46079c AK |
893 | } |
894 | ||
895 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 896 | unsigned long pfn) |
6a46079c AK |
897 | { |
898 | int result; | |
7456b040 | 899 | int count; |
6a46079c AK |
900 | |
901 | result = ps->action(p, pfn); | |
7456b040 | 902 | |
bd1ce5f9 | 903 | count = page_count(p) - 1; |
cc637b17 | 904 | if (ps->action == me_swapcache_dirty && result == MF_DELAYED) |
138ce286 | 905 | count--; |
78bb9203 | 906 | if (count > 0) { |
495367c0 | 907 | pr_err("Memory failure: %#lx: %s still referenced by %d users\n", |
64d37a2b | 908 | pfn, action_page_types[ps->type], count); |
cc637b17 | 909 | result = MF_FAILED; |
138ce286 | 910 | } |
64d37a2b | 911 | action_result(pfn, ps->type, result); |
6a46079c AK |
912 | |
913 | /* Could do more checks here if page looks ok */ | |
914 | /* | |
915 | * Could adjust zone counters here to correct for the missing page. | |
916 | */ | |
917 | ||
cc637b17 | 918 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
919 | } |
920 | ||
ead07f6a NH |
921 | /** |
922 | * get_hwpoison_page() - Get refcount for memory error handling: | |
923 | * @page: raw error page (hit by memory error) | |
924 | * | |
925 | * Return: return 0 if failed to grab the refcount, otherwise true (some | |
926 | * non-zero value.) | |
927 | */ | |
928 | int get_hwpoison_page(struct page *page) | |
929 | { | |
930 | struct page *head = compound_head(page); | |
931 | ||
4e41a30c | 932 | if (!PageHuge(head) && PageTransHuge(head)) { |
98ed2b00 NH |
933 | /* |
934 | * Non anonymous thp exists only in allocation/free time. We | |
935 | * can't handle such a case correctly, so let's give it up. | |
936 | * This should be better than triggering BUG_ON when kernel | |
937 | * tries to touch the "partially handled" page. | |
938 | */ | |
939 | if (!PageAnon(head)) { | |
495367c0 | 940 | pr_err("Memory failure: %#lx: non anonymous thp\n", |
98ed2b00 NH |
941 | page_to_pfn(page)); |
942 | return 0; | |
943 | } | |
ead07f6a NH |
944 | } |
945 | ||
c2e7e00b KK |
946 | if (get_page_unless_zero(head)) { |
947 | if (head == compound_head(page)) | |
948 | return 1; | |
949 | ||
495367c0 CY |
950 | pr_info("Memory failure: %#lx cannot catch tail\n", |
951 | page_to_pfn(page)); | |
c2e7e00b KK |
952 | put_page(head); |
953 | } | |
954 | ||
955 | return 0; | |
ead07f6a NH |
956 | } |
957 | EXPORT_SYMBOL_GPL(get_hwpoison_page); | |
958 | ||
6a46079c AK |
959 | /* |
960 | * Do all that is necessary to remove user space mappings. Unmap | |
961 | * the pages and send SIGBUS to the processes if the data was dirty. | |
962 | */ | |
666e5a40 | 963 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
83b57531 | 964 | int flags, struct page **hpagep) |
6a46079c | 965 | { |
a128ca71 | 966 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; |
6a46079c AK |
967 | struct address_space *mapping; |
968 | LIST_HEAD(tokill); | |
666e5a40 | 969 | bool unmap_success; |
6751ed65 | 970 | int kill = 1, forcekill; |
54b9dd14 | 971 | struct page *hpage = *hpagep; |
286c469a | 972 | bool mlocked = PageMlocked(hpage); |
6a46079c | 973 | |
93a9eb39 NH |
974 | /* |
975 | * Here we are interested only in user-mapped pages, so skip any | |
976 | * other types of pages. | |
977 | */ | |
978 | if (PageReserved(p) || PageSlab(p)) | |
666e5a40 | 979 | return true; |
93a9eb39 | 980 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 981 | return true; |
6a46079c | 982 | |
6a46079c AK |
983 | /* |
984 | * This check implies we don't kill processes if their pages | |
985 | * are in the swap cache early. Those are always late kills. | |
986 | */ | |
7af446a8 | 987 | if (!page_mapped(hpage)) |
666e5a40 | 988 | return true; |
1668bfd5 | 989 | |
52089b14 | 990 | if (PageKsm(p)) { |
495367c0 | 991 | pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn); |
666e5a40 | 992 | return false; |
52089b14 | 993 | } |
6a46079c AK |
994 | |
995 | if (PageSwapCache(p)) { | |
495367c0 CY |
996 | pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n", |
997 | pfn); | |
6a46079c AK |
998 | ttu |= TTU_IGNORE_HWPOISON; |
999 | } | |
1000 | ||
1001 | /* | |
1002 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1003 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1004 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1005 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1006 | */ |
7af446a8 | 1007 | mapping = page_mapping(hpage); |
6751ed65 | 1008 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
7af446a8 NH |
1009 | mapping_cap_writeback_dirty(mapping)) { |
1010 | if (page_mkclean(hpage)) { | |
1011 | SetPageDirty(hpage); | |
6a46079c AK |
1012 | } else { |
1013 | kill = 0; | |
1014 | ttu |= TTU_IGNORE_HWPOISON; | |
495367c0 | 1015 | pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1016 | pfn); |
1017 | } | |
1018 | } | |
1019 | ||
1020 | /* | |
1021 | * First collect all the processes that have the page | |
1022 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1023 | * because ttu takes the rmap data structures down. | |
1024 | * | |
1025 | * Error handling: We ignore errors here because | |
1026 | * there's nothing that can be done. | |
1027 | */ | |
1028 | if (kill) | |
415c64c1 | 1029 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1030 | |
666e5a40 MK |
1031 | unmap_success = try_to_unmap(hpage, ttu); |
1032 | if (!unmap_success) | |
495367c0 | 1033 | pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1034 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1035 | |
286c469a NH |
1036 | /* |
1037 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1038 | * shake_page() again to ensure that it's flushed. | |
1039 | */ | |
1040 | if (mlocked) | |
1041 | shake_page(hpage, 0); | |
1042 | ||
6a46079c AK |
1043 | /* |
1044 | * Now that the dirty bit has been propagated to the | |
1045 | * struct page and all unmaps done we can decide if | |
1046 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1047 | * was dirty or the process is not restartable, |
1048 | * otherwise the tokill list is merely | |
6a46079c AK |
1049 | * freed. When there was a problem unmapping earlier |
1050 | * use a more force-full uncatchable kill to prevent | |
1051 | * any accesses to the poisoned memory. | |
1052 | */ | |
415c64c1 | 1053 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
ae1139ec | 1054 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1055 | |
666e5a40 | 1056 | return unmap_success; |
6a46079c AK |
1057 | } |
1058 | ||
0348d2eb NH |
1059 | static int identify_page_state(unsigned long pfn, struct page *p, |
1060 | unsigned long page_flags) | |
761ad8d7 NH |
1061 | { |
1062 | struct page_state *ps; | |
0348d2eb NH |
1063 | |
1064 | /* | |
1065 | * The first check uses the current page flags which may not have any | |
1066 | * relevant information. The second check with the saved page flags is | |
1067 | * carried out only if the first check can't determine the page status. | |
1068 | */ | |
1069 | for (ps = error_states;; ps++) | |
1070 | if ((p->flags & ps->mask) == ps->res) | |
1071 | break; | |
1072 | ||
1073 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1074 | ||
1075 | if (!ps->mask) | |
1076 | for (ps = error_states;; ps++) | |
1077 | if ((page_flags & ps->mask) == ps->res) | |
1078 | break; | |
1079 | return page_action(ps, p, pfn); | |
1080 | } | |
1081 | ||
83b57531 | 1082 | static int memory_failure_hugetlb(unsigned long pfn, int flags) |
0348d2eb | 1083 | { |
761ad8d7 NH |
1084 | struct page *p = pfn_to_page(pfn); |
1085 | struct page *head = compound_head(p); | |
1086 | int res; | |
1087 | unsigned long page_flags; | |
1088 | ||
1089 | if (TestSetPageHWPoison(head)) { | |
1090 | pr_err("Memory failure: %#lx: already hardware poisoned\n", | |
1091 | pfn); | |
1092 | return 0; | |
1093 | } | |
1094 | ||
1095 | num_poisoned_pages_inc(); | |
1096 | ||
1097 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { | |
1098 | /* | |
1099 | * Check "filter hit" and "race with other subpage." | |
1100 | */ | |
1101 | lock_page(head); | |
1102 | if (PageHWPoison(head)) { | |
1103 | if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1104 | || (p != head && TestSetPageHWPoison(head))) { | |
1105 | num_poisoned_pages_dec(); | |
1106 | unlock_page(head); | |
1107 | return 0; | |
1108 | } | |
1109 | } | |
1110 | unlock_page(head); | |
1111 | dissolve_free_huge_page(p); | |
1112 | action_result(pfn, MF_MSG_FREE_HUGE, MF_DELAYED); | |
1113 | return 0; | |
1114 | } | |
1115 | ||
1116 | lock_page(head); | |
1117 | page_flags = head->flags; | |
1118 | ||
1119 | if (!PageHWPoison(head)) { | |
1120 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); | |
1121 | num_poisoned_pages_dec(); | |
1122 | unlock_page(head); | |
1123 | put_hwpoison_page(head); | |
1124 | return 0; | |
1125 | } | |
1126 | ||
31286a84 NH |
1127 | /* |
1128 | * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so | |
1129 | * simply disable it. In order to make it work properly, we need | |
1130 | * make sure that: | |
1131 | * - conversion of a pud that maps an error hugetlb into hwpoison | |
1132 | * entry properly works, and | |
1133 | * - other mm code walking over page table is aware of pud-aligned | |
1134 | * hwpoison entries. | |
1135 | */ | |
1136 | if (huge_page_size(page_hstate(head)) > PMD_SIZE) { | |
1137 | action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED); | |
1138 | res = -EBUSY; | |
1139 | goto out; | |
1140 | } | |
1141 | ||
83b57531 | 1142 | if (!hwpoison_user_mappings(p, pfn, flags, &head)) { |
761ad8d7 NH |
1143 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1144 | res = -EBUSY; | |
1145 | goto out; | |
1146 | } | |
1147 | ||
0348d2eb | 1148 | res = identify_page_state(pfn, p, page_flags); |
761ad8d7 NH |
1149 | out: |
1150 | unlock_page(head); | |
1151 | return res; | |
1152 | } | |
1153 | ||
6100e34b DW |
1154 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
1155 | struct dev_pagemap *pgmap) | |
1156 | { | |
1157 | struct page *page = pfn_to_page(pfn); | |
1158 | const bool unmap_success = true; | |
1159 | unsigned long size = 0; | |
1160 | struct to_kill *tk; | |
1161 | LIST_HEAD(tokill); | |
1162 | int rc = -EBUSY; | |
1163 | loff_t start; | |
1164 | ||
1165 | /* | |
1166 | * Prevent the inode from being freed while we are interrogating | |
1167 | * the address_space, typically this would be handled by | |
1168 | * lock_page(), but dax pages do not use the page lock. This | |
1169 | * also prevents changes to the mapping of this pfn until | |
1170 | * poison signaling is complete. | |
1171 | */ | |
1172 | if (!dax_lock_mapping_entry(page)) | |
1173 | goto out; | |
1174 | ||
1175 | if (hwpoison_filter(page)) { | |
1176 | rc = 0; | |
1177 | goto unlock; | |
1178 | } | |
1179 | ||
1180 | switch (pgmap->type) { | |
1181 | case MEMORY_DEVICE_PRIVATE: | |
1182 | case MEMORY_DEVICE_PUBLIC: | |
1183 | /* | |
1184 | * TODO: Handle HMM pages which may need coordination | |
1185 | * with device-side memory. | |
1186 | */ | |
1187 | goto unlock; | |
1188 | default: | |
1189 | break; | |
1190 | } | |
1191 | ||
1192 | /* | |
1193 | * Use this flag as an indication that the dax page has been | |
1194 | * remapped UC to prevent speculative consumption of poison. | |
1195 | */ | |
1196 | SetPageHWPoison(page); | |
1197 | ||
1198 | /* | |
1199 | * Unlike System-RAM there is no possibility to swap in a | |
1200 | * different physical page at a given virtual address, so all | |
1201 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1202 | * SIGBUS (i.e. MF_MUST_KILL) | |
1203 | */ | |
1204 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1205 | collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED); | |
1206 | ||
1207 | list_for_each_entry(tk, &tokill, nd) | |
1208 | if (tk->size_shift) | |
1209 | size = max(size, 1UL << tk->size_shift); | |
1210 | if (size) { | |
1211 | /* | |
1212 | * Unmap the largest mapping to avoid breaking up | |
1213 | * device-dax mappings which are constant size. The | |
1214 | * actual size of the mapping being torn down is | |
1215 | * communicated in siginfo, see kill_proc() | |
1216 | */ | |
1217 | start = (page->index << PAGE_SHIFT) & ~(size - 1); | |
1218 | unmap_mapping_range(page->mapping, start, start + size, 0); | |
1219 | } | |
1220 | kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags); | |
1221 | rc = 0; | |
1222 | unlock: | |
1223 | dax_unlock_mapping_entry(page); | |
1224 | out: | |
1225 | /* drop pgmap ref acquired in caller */ | |
1226 | put_dev_pagemap(pgmap); | |
1227 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
1228 | return rc; | |
1229 | } | |
1230 | ||
cd42f4a3 TL |
1231 | /** |
1232 | * memory_failure - Handle memory failure of a page. | |
1233 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
1234 | * @flags: fine tune action taken |
1235 | * | |
1236 | * This function is called by the low level machine check code | |
1237 | * of an architecture when it detects hardware memory corruption | |
1238 | * of a page. It tries its best to recover, which includes | |
1239 | * dropping pages, killing processes etc. | |
1240 | * | |
1241 | * The function is primarily of use for corruptions that | |
1242 | * happen outside the current execution context (e.g. when | |
1243 | * detected by a background scrubber) | |
1244 | * | |
1245 | * Must run in process context (e.g. a work queue) with interrupts | |
1246 | * enabled and no spinlocks hold. | |
1247 | */ | |
83b57531 | 1248 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 1249 | { |
6a46079c | 1250 | struct page *p; |
7af446a8 | 1251 | struct page *hpage; |
415c64c1 | 1252 | struct page *orig_head; |
6100e34b | 1253 | struct dev_pagemap *pgmap; |
6a46079c | 1254 | int res; |
524fca1e | 1255 | unsigned long page_flags; |
6a46079c AK |
1256 | |
1257 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 1258 | panic("Memory failure on page %lx", pfn); |
6a46079c AK |
1259 | |
1260 | if (!pfn_valid(pfn)) { | |
495367c0 CY |
1261 | pr_err("Memory failure: %#lx: memory outside kernel control\n", |
1262 | pfn); | |
a7560fc8 | 1263 | return -ENXIO; |
6a46079c AK |
1264 | } |
1265 | ||
6100e34b DW |
1266 | pgmap = get_dev_pagemap(pfn, NULL); |
1267 | if (pgmap) | |
1268 | return memory_failure_dev_pagemap(pfn, flags, pgmap); | |
1269 | ||
6a46079c | 1270 | p = pfn_to_page(pfn); |
761ad8d7 | 1271 | if (PageHuge(p)) |
83b57531 | 1272 | return memory_failure_hugetlb(pfn, flags); |
6a46079c | 1273 | if (TestSetPageHWPoison(p)) { |
495367c0 CY |
1274 | pr_err("Memory failure: %#lx: already hardware poisoned\n", |
1275 | pfn); | |
6a46079c AK |
1276 | return 0; |
1277 | } | |
1278 | ||
761ad8d7 | 1279 | orig_head = hpage = compound_head(p); |
b37ff71c | 1280 | num_poisoned_pages_inc(); |
6a46079c AK |
1281 | |
1282 | /* | |
1283 | * We need/can do nothing about count=0 pages. | |
1284 | * 1) it's a free page, and therefore in safe hand: | |
1285 | * prep_new_page() will be the gate keeper. | |
761ad8d7 | 1286 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
1287 | * Implies some kernel user: cannot stop them from |
1288 | * R/W the page; let's pray that the page has been | |
1289 | * used and will be freed some time later. | |
1290 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 1291 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 1292 | */ |
ead07f6a | 1293 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { |
8d22ba1b | 1294 | if (is_free_buddy_page(p)) { |
cc637b17 | 1295 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); |
8d22ba1b WF |
1296 | return 0; |
1297 | } else { | |
cc637b17 | 1298 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); |
8d22ba1b WF |
1299 | return -EBUSY; |
1300 | } | |
6a46079c AK |
1301 | } |
1302 | ||
761ad8d7 | 1303 | if (PageTransHuge(hpage)) { |
c3901e72 NH |
1304 | lock_page(p); |
1305 | if (!PageAnon(p) || unlikely(split_huge_page(p))) { | |
1306 | unlock_page(p); | |
1307 | if (!PageAnon(p)) | |
495367c0 CY |
1308 | pr_err("Memory failure: %#lx: non anonymous thp\n", |
1309 | pfn); | |
7f6bf39b | 1310 | else |
495367c0 CY |
1311 | pr_err("Memory failure: %#lx: thp split failed\n", |
1312 | pfn); | |
ead07f6a | 1313 | if (TestClearPageHWPoison(p)) |
b37ff71c | 1314 | num_poisoned_pages_dec(); |
665d9da7 | 1315 | put_hwpoison_page(p); |
415c64c1 NH |
1316 | return -EBUSY; |
1317 | } | |
c3901e72 | 1318 | unlock_page(p); |
415c64c1 NH |
1319 | VM_BUG_ON_PAGE(!page_count(p), p); |
1320 | hpage = compound_head(p); | |
1321 | } | |
1322 | ||
e43c3afb WF |
1323 | /* |
1324 | * We ignore non-LRU pages for good reasons. | |
1325 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 1326 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
1327 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
1328 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1329 | * walked by the page reclaim code, however that's not a big loss. | |
1330 | */ | |
8bcb74de NH |
1331 | shake_page(p, 0); |
1332 | /* shake_page could have turned it free. */ | |
1333 | if (!PageLRU(p) && is_free_buddy_page(p)) { | |
1334 | if (flags & MF_COUNT_INCREASED) | |
1335 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); | |
1336 | else | |
1337 | action_result(pfn, MF_MSG_BUDDY_2ND, MF_DELAYED); | |
1338 | return 0; | |
e43c3afb | 1339 | } |
e43c3afb | 1340 | |
761ad8d7 | 1341 | lock_page(p); |
847ce401 | 1342 | |
f37d4298 AK |
1343 | /* |
1344 | * The page could have changed compound pages during the locking. | |
1345 | * If this happens just bail out. | |
1346 | */ | |
415c64c1 | 1347 | if (PageCompound(p) && compound_head(p) != orig_head) { |
cc637b17 | 1348 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 AK |
1349 | res = -EBUSY; |
1350 | goto out; | |
1351 | } | |
1352 | ||
524fca1e NH |
1353 | /* |
1354 | * We use page flags to determine what action should be taken, but | |
1355 | * the flags can be modified by the error containment action. One | |
1356 | * example is an mlocked page, where PG_mlocked is cleared by | |
1357 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1358 | * correctly, we save a copy of the page flags at this time. | |
1359 | */ | |
7258ae5c JM |
1360 | if (PageHuge(p)) |
1361 | page_flags = hpage->flags; | |
1362 | else | |
1363 | page_flags = p->flags; | |
524fca1e | 1364 | |
847ce401 WF |
1365 | /* |
1366 | * unpoison always clear PG_hwpoison inside page lock | |
1367 | */ | |
1368 | if (!PageHWPoison(p)) { | |
495367c0 | 1369 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); |
b37ff71c | 1370 | num_poisoned_pages_dec(); |
761ad8d7 NH |
1371 | unlock_page(p); |
1372 | put_hwpoison_page(p); | |
a09233f3 | 1373 | return 0; |
847ce401 | 1374 | } |
7c116f2b WF |
1375 | if (hwpoison_filter(p)) { |
1376 | if (TestClearPageHWPoison(p)) | |
b37ff71c | 1377 | num_poisoned_pages_dec(); |
761ad8d7 NH |
1378 | unlock_page(p); |
1379 | put_hwpoison_page(p); | |
7c116f2b WF |
1380 | return 0; |
1381 | } | |
847ce401 | 1382 | |
761ad8d7 | 1383 | if (!PageTransTail(p) && !PageLRU(p)) |
0bc1f8b0 CY |
1384 | goto identify_page_state; |
1385 | ||
6edd6cc6 NH |
1386 | /* |
1387 | * It's very difficult to mess with pages currently under IO | |
1388 | * and in many cases impossible, so we just avoid it here. | |
1389 | */ | |
6a46079c AK |
1390 | wait_on_page_writeback(p); |
1391 | ||
1392 | /* | |
1393 | * Now take care of user space mappings. | |
e64a782f | 1394 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
54b9dd14 NH |
1395 | * |
1396 | * When the raw error page is thp tail page, hpage points to the raw | |
1397 | * page after thp split. | |
6a46079c | 1398 | */ |
83b57531 | 1399 | if (!hwpoison_user_mappings(p, pfn, flags, &hpage)) { |
cc637b17 | 1400 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 WF |
1401 | res = -EBUSY; |
1402 | goto out; | |
1403 | } | |
6a46079c AK |
1404 | |
1405 | /* | |
1406 | * Torn down by someone else? | |
1407 | */ | |
dc2a1cbf | 1408 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 1409 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 1410 | res = -EBUSY; |
6a46079c AK |
1411 | goto out; |
1412 | } | |
1413 | ||
0bc1f8b0 | 1414 | identify_page_state: |
0348d2eb | 1415 | res = identify_page_state(pfn, p, page_flags); |
6a46079c | 1416 | out: |
761ad8d7 | 1417 | unlock_page(p); |
6a46079c AK |
1418 | return res; |
1419 | } | |
cd42f4a3 | 1420 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1421 | |
ea8f5fb8 HY |
1422 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1423 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1424 | ||
1425 | struct memory_failure_entry { | |
1426 | unsigned long pfn; | |
ea8f5fb8 HY |
1427 | int flags; |
1428 | }; | |
1429 | ||
1430 | struct memory_failure_cpu { | |
1431 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1432 | MEMORY_FAILURE_FIFO_SIZE); | |
1433 | spinlock_t lock; | |
1434 | struct work_struct work; | |
1435 | }; | |
1436 | ||
1437 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1438 | ||
1439 | /** | |
1440 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1441 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
1442 | * @flags: Flags for memory failure handling |
1443 | * | |
1444 | * This function is called by the low level hardware error handler | |
1445 | * when it detects hardware memory corruption of a page. It schedules | |
1446 | * the recovering of error page, including dropping pages, killing | |
1447 | * processes etc. | |
1448 | * | |
1449 | * The function is primarily of use for corruptions that | |
1450 | * happen outside the current execution context (e.g. when | |
1451 | * detected by a background scrubber) | |
1452 | * | |
1453 | * Can run in IRQ context. | |
1454 | */ | |
83b57531 | 1455 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
1456 | { |
1457 | struct memory_failure_cpu *mf_cpu; | |
1458 | unsigned long proc_flags; | |
1459 | struct memory_failure_entry entry = { | |
1460 | .pfn = pfn, | |
ea8f5fb8 HY |
1461 | .flags = flags, |
1462 | }; | |
1463 | ||
1464 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1465 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 1466 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
1467 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
1468 | else | |
8e33a52f | 1469 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
1470 | pfn); |
1471 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1472 | put_cpu_var(memory_failure_cpu); | |
1473 | } | |
1474 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1475 | ||
1476 | static void memory_failure_work_func(struct work_struct *work) | |
1477 | { | |
1478 | struct memory_failure_cpu *mf_cpu; | |
1479 | struct memory_failure_entry entry = { 0, }; | |
1480 | unsigned long proc_flags; | |
1481 | int gotten; | |
1482 | ||
7c8e0181 | 1483 | mf_cpu = this_cpu_ptr(&memory_failure_cpu); |
ea8f5fb8 HY |
1484 | for (;;) { |
1485 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1486 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1487 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1488 | if (!gotten) | |
1489 | break; | |
cf870c70 NR |
1490 | if (entry.flags & MF_SOFT_OFFLINE) |
1491 | soft_offline_page(pfn_to_page(entry.pfn), entry.flags); | |
1492 | else | |
83b57531 | 1493 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
1494 | } |
1495 | } | |
1496 | ||
1497 | static int __init memory_failure_init(void) | |
1498 | { | |
1499 | struct memory_failure_cpu *mf_cpu; | |
1500 | int cpu; | |
1501 | ||
1502 | for_each_possible_cpu(cpu) { | |
1503 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1504 | spin_lock_init(&mf_cpu->lock); | |
1505 | INIT_KFIFO(mf_cpu->fifo); | |
1506 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1507 | } | |
1508 | ||
1509 | return 0; | |
1510 | } | |
1511 | core_initcall(memory_failure_init); | |
1512 | ||
a5f65109 NH |
1513 | #define unpoison_pr_info(fmt, pfn, rs) \ |
1514 | ({ \ | |
1515 | if (__ratelimit(rs)) \ | |
1516 | pr_info(fmt, pfn); \ | |
1517 | }) | |
1518 | ||
847ce401 WF |
1519 | /** |
1520 | * unpoison_memory - Unpoison a previously poisoned page | |
1521 | * @pfn: Page number of the to be unpoisoned page | |
1522 | * | |
1523 | * Software-unpoison a page that has been poisoned by | |
1524 | * memory_failure() earlier. | |
1525 | * | |
1526 | * This is only done on the software-level, so it only works | |
1527 | * for linux injected failures, not real hardware failures | |
1528 | * | |
1529 | * Returns 0 for success, otherwise -errno. | |
1530 | */ | |
1531 | int unpoison_memory(unsigned long pfn) | |
1532 | { | |
1533 | struct page *page; | |
1534 | struct page *p; | |
1535 | int freeit = 0; | |
a5f65109 NH |
1536 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
1537 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
1538 | |
1539 | if (!pfn_valid(pfn)) | |
1540 | return -ENXIO; | |
1541 | ||
1542 | p = pfn_to_page(pfn); | |
1543 | page = compound_head(p); | |
1544 | ||
1545 | if (!PageHWPoison(p)) { | |
495367c0 | 1546 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 1547 | pfn, &unpoison_rs); |
847ce401 WF |
1548 | return 0; |
1549 | } | |
1550 | ||
230ac719 | 1551 | if (page_count(page) > 1) { |
495367c0 | 1552 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 1553 | pfn, &unpoison_rs); |
230ac719 NH |
1554 | return 0; |
1555 | } | |
1556 | ||
1557 | if (page_mapped(page)) { | |
495367c0 | 1558 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 1559 | pfn, &unpoison_rs); |
230ac719 NH |
1560 | return 0; |
1561 | } | |
1562 | ||
1563 | if (page_mapping(page)) { | |
495367c0 | 1564 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 1565 | pfn, &unpoison_rs); |
230ac719 NH |
1566 | return 0; |
1567 | } | |
1568 | ||
0cea3fdc WL |
1569 | /* |
1570 | * unpoison_memory() can encounter thp only when the thp is being | |
1571 | * worked by memory_failure() and the page lock is not held yet. | |
1572 | * In such case, we yield to memory_failure() and make unpoison fail. | |
1573 | */ | |
e76d30e2 | 1574 | if (!PageHuge(page) && PageTransHuge(page)) { |
495367c0 | 1575 | unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n", |
a5f65109 | 1576 | pfn, &unpoison_rs); |
ead07f6a | 1577 | return 0; |
0cea3fdc WL |
1578 | } |
1579 | ||
ead07f6a | 1580 | if (!get_hwpoison_page(p)) { |
847ce401 | 1581 | if (TestClearPageHWPoison(p)) |
8e30456b | 1582 | num_poisoned_pages_dec(); |
495367c0 | 1583 | unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n", |
a5f65109 | 1584 | pfn, &unpoison_rs); |
847ce401 WF |
1585 | return 0; |
1586 | } | |
1587 | ||
7eaceacc | 1588 | lock_page(page); |
847ce401 WF |
1589 | /* |
1590 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1591 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1592 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1593 | * the free buddy page pool. | |
1594 | */ | |
c9fbdd5f | 1595 | if (TestClearPageHWPoison(page)) { |
495367c0 | 1596 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", |
a5f65109 | 1597 | pfn, &unpoison_rs); |
b37ff71c | 1598 | num_poisoned_pages_dec(); |
847ce401 WF |
1599 | freeit = 1; |
1600 | } | |
1601 | unlock_page(page); | |
1602 | ||
665d9da7 | 1603 | put_hwpoison_page(page); |
3ba5eebc | 1604 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) |
665d9da7 | 1605 | put_hwpoison_page(page); |
847ce401 WF |
1606 | |
1607 | return 0; | |
1608 | } | |
1609 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 1610 | |
666feb21 | 1611 | static struct page *new_page(struct page *p, unsigned long private) |
facb6011 | 1612 | { |
12686d15 | 1613 | int nid = page_to_nid(p); |
94310cbc | 1614 | |
ef77ba5c | 1615 | return new_page_nodemask(p, nid, &node_states[N_MEMORY]); |
facb6011 AK |
1616 | } |
1617 | ||
1618 | /* | |
1619 | * Safely get reference count of an arbitrary page. | |
1620 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1621 | * that is not free, and 1 for any other page type. | |
1622 | * For 1 the page is returned with increased page count, otherwise not. | |
1623 | */ | |
af8fae7c | 1624 | static int __get_any_page(struct page *p, unsigned long pfn, int flags) |
facb6011 AK |
1625 | { |
1626 | int ret; | |
1627 | ||
1628 | if (flags & MF_COUNT_INCREASED) | |
1629 | return 1; | |
1630 | ||
d950b958 NH |
1631 | /* |
1632 | * When the target page is a free hugepage, just remove it | |
1633 | * from free hugepage list. | |
1634 | */ | |
ead07f6a | 1635 | if (!get_hwpoison_page(p)) { |
d950b958 | 1636 | if (PageHuge(p)) { |
71dd0b8a | 1637 | pr_info("%s: %#lx free huge page\n", __func__, pfn); |
af8fae7c | 1638 | ret = 0; |
d950b958 | 1639 | } else if (is_free_buddy_page(p)) { |
71dd0b8a | 1640 | pr_info("%s: %#lx free buddy page\n", __func__, pfn); |
facb6011 AK |
1641 | ret = 0; |
1642 | } else { | |
71dd0b8a BP |
1643 | pr_info("%s: %#lx: unknown zero refcount page type %lx\n", |
1644 | __func__, pfn, p->flags); | |
facb6011 AK |
1645 | ret = -EIO; |
1646 | } | |
1647 | } else { | |
1648 | /* Not a free page */ | |
1649 | ret = 1; | |
1650 | } | |
facb6011 AK |
1651 | return ret; |
1652 | } | |
1653 | ||
af8fae7c NH |
1654 | static int get_any_page(struct page *page, unsigned long pfn, int flags) |
1655 | { | |
1656 | int ret = __get_any_page(page, pfn, flags); | |
1657 | ||
85fbe5d1 YX |
1658 | if (ret == 1 && !PageHuge(page) && |
1659 | !PageLRU(page) && !__PageMovable(page)) { | |
af8fae7c NH |
1660 | /* |
1661 | * Try to free it. | |
1662 | */ | |
665d9da7 | 1663 | put_hwpoison_page(page); |
af8fae7c NH |
1664 | shake_page(page, 1); |
1665 | ||
1666 | /* | |
1667 | * Did it turn free? | |
1668 | */ | |
1669 | ret = __get_any_page(page, pfn, 0); | |
d96b339f | 1670 | if (ret == 1 && !PageLRU(page)) { |
4f32be67 | 1671 | /* Drop page reference which is from __get_any_page() */ |
665d9da7 | 1672 | put_hwpoison_page(page); |
82a2481e AK |
1673 | pr_info("soft_offline: %#lx: unknown non LRU page type %lx (%pGp)\n", |
1674 | pfn, page->flags, &page->flags); | |
af8fae7c NH |
1675 | return -EIO; |
1676 | } | |
1677 | } | |
1678 | return ret; | |
1679 | } | |
1680 | ||
d950b958 NH |
1681 | static int soft_offline_huge_page(struct page *page, int flags) |
1682 | { | |
1683 | int ret; | |
1684 | unsigned long pfn = page_to_pfn(page); | |
1685 | struct page *hpage = compound_head(page); | |
b8ec1cee | 1686 | LIST_HEAD(pagelist); |
d950b958 | 1687 | |
af8fae7c NH |
1688 | /* |
1689 | * This double-check of PageHWPoison is to avoid the race with | |
1690 | * memory_failure(). See also comment in __soft_offline_page(). | |
1691 | */ | |
1692 | lock_page(hpage); | |
0ebff32c | 1693 | if (PageHWPoison(hpage)) { |
af8fae7c | 1694 | unlock_page(hpage); |
665d9da7 | 1695 | put_hwpoison_page(hpage); |
0ebff32c | 1696 | pr_info("soft offline: %#lx hugepage already poisoned\n", pfn); |
af8fae7c | 1697 | return -EBUSY; |
0ebff32c | 1698 | } |
af8fae7c | 1699 | unlock_page(hpage); |
d950b958 | 1700 | |
bcc54222 | 1701 | ret = isolate_huge_page(hpage, &pagelist); |
03613808 WL |
1702 | /* |
1703 | * get_any_page() and isolate_huge_page() takes a refcount each, | |
1704 | * so need to drop one here. | |
1705 | */ | |
665d9da7 | 1706 | put_hwpoison_page(hpage); |
03613808 | 1707 | if (!ret) { |
bcc54222 NH |
1708 | pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn); |
1709 | return -EBUSY; | |
1710 | } | |
1711 | ||
68711a74 | 1712 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, |
b8ec1cee | 1713 | MIGRATE_SYNC, MR_MEMORY_FAILURE); |
d950b958 | 1714 | if (ret) { |
b6b18aa8 | 1715 | pr_info("soft offline: %#lx: hugepage migration failed %d, type %lx (%pGp)\n", |
82a2481e | 1716 | pfn, ret, page->flags, &page->flags); |
30809f55 PA |
1717 | if (!list_empty(&pagelist)) |
1718 | putback_movable_pages(&pagelist); | |
b8ec1cee NH |
1719 | if (ret > 0) |
1720 | ret = -EIO; | |
af8fae7c | 1721 | } else { |
6bc9b564 NH |
1722 | /* |
1723 | * We set PG_hwpoison only when the migration source hugepage | |
1724 | * was successfully dissolved, because otherwise hwpoisoned | |
1725 | * hugepage remains on free hugepage list, then userspace will | |
1726 | * find it as SIGBUS by allocation failure. That's not expected | |
1727 | * in soft-offlining. | |
1728 | */ | |
1729 | ret = dissolve_free_huge_page(page); | |
1730 | if (!ret) { | |
1731 | if (set_hwpoison_free_buddy_page(page)) | |
1732 | num_poisoned_pages_inc(); | |
1733 | } | |
d950b958 | 1734 | } |
d950b958 NH |
1735 | return ret; |
1736 | } | |
1737 | ||
af8fae7c NH |
1738 | static int __soft_offline_page(struct page *page, int flags) |
1739 | { | |
1740 | int ret; | |
1741 | unsigned long pfn = page_to_pfn(page); | |
facb6011 | 1742 | |
facb6011 | 1743 | /* |
af8fae7c NH |
1744 | * Check PageHWPoison again inside page lock because PageHWPoison |
1745 | * is set by memory_failure() outside page lock. Note that | |
1746 | * memory_failure() also double-checks PageHWPoison inside page lock, | |
1747 | * so there's no race between soft_offline_page() and memory_failure(). | |
facb6011 | 1748 | */ |
0ebff32c XQ |
1749 | lock_page(page); |
1750 | wait_on_page_writeback(page); | |
af8fae7c NH |
1751 | if (PageHWPoison(page)) { |
1752 | unlock_page(page); | |
665d9da7 | 1753 | put_hwpoison_page(page); |
af8fae7c NH |
1754 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
1755 | return -EBUSY; | |
1756 | } | |
facb6011 AK |
1757 | /* |
1758 | * Try to invalidate first. This should work for | |
1759 | * non dirty unmapped page cache pages. | |
1760 | */ | |
1761 | ret = invalidate_inode_page(page); | |
1762 | unlock_page(page); | |
facb6011 | 1763 | /* |
facb6011 AK |
1764 | * RED-PEN would be better to keep it isolated here, but we |
1765 | * would need to fix isolation locking first. | |
1766 | */ | |
facb6011 | 1767 | if (ret == 1) { |
665d9da7 | 1768 | put_hwpoison_page(page); |
fb46e735 | 1769 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
af8fae7c | 1770 | SetPageHWPoison(page); |
8e30456b | 1771 | num_poisoned_pages_inc(); |
af8fae7c | 1772 | return 0; |
facb6011 AK |
1773 | } |
1774 | ||
1775 | /* | |
1776 | * Simple invalidation didn't work. | |
1777 | * Try to migrate to a new page instead. migrate.c | |
1778 | * handles a large number of cases for us. | |
1779 | */ | |
85fbe5d1 YX |
1780 | if (PageLRU(page)) |
1781 | ret = isolate_lru_page(page); | |
1782 | else | |
1783 | ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); | |
bd486285 KK |
1784 | /* |
1785 | * Drop page reference which is came from get_any_page() | |
1786 | * successful isolate_lru_page() already took another one. | |
1787 | */ | |
665d9da7 | 1788 | put_hwpoison_page(page); |
facb6011 AK |
1789 | if (!ret) { |
1790 | LIST_HEAD(pagelist); | |
85fbe5d1 YX |
1791 | /* |
1792 | * After isolated lru page, the PageLRU will be cleared, | |
1793 | * so use !__PageMovable instead for LRU page's mapping | |
1794 | * cannot have PAGE_MAPPING_MOVABLE. | |
1795 | */ | |
1796 | if (!__PageMovable(page)) | |
1797 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
1798 | page_is_file_cache(page)); | |
facb6011 | 1799 | list_add(&page->lru, &pagelist); |
68711a74 | 1800 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, |
9c620e2b | 1801 | MIGRATE_SYNC, MR_MEMORY_FAILURE); |
facb6011 | 1802 | if (ret) { |
85fbe5d1 YX |
1803 | if (!list_empty(&pagelist)) |
1804 | putback_movable_pages(&pagelist); | |
59c82b70 | 1805 | |
82a2481e AK |
1806 | pr_info("soft offline: %#lx: migration failed %d, type %lx (%pGp)\n", |
1807 | pfn, ret, page->flags, &page->flags); | |
facb6011 AK |
1808 | if (ret > 0) |
1809 | ret = -EIO; | |
1810 | } | |
1811 | } else { | |
82a2481e AK |
1812 | pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx (%pGp)\n", |
1813 | pfn, ret, page_count(page), page->flags, &page->flags); | |
facb6011 | 1814 | } |
facb6011 AK |
1815 | return ret; |
1816 | } | |
86e05773 | 1817 | |
acc14dc4 NH |
1818 | static int soft_offline_in_use_page(struct page *page, int flags) |
1819 | { | |
1820 | int ret; | |
d4ae9916 | 1821 | int mt; |
acc14dc4 NH |
1822 | struct page *hpage = compound_head(page); |
1823 | ||
1824 | if (!PageHuge(page) && PageTransHuge(hpage)) { | |
1825 | lock_page(hpage); | |
98fd1ef4 NH |
1826 | if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) { |
1827 | unlock_page(hpage); | |
1828 | if (!PageAnon(hpage)) | |
1829 | pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page)); | |
1830 | else | |
1831 | pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page)); | |
1832 | put_hwpoison_page(hpage); | |
acc14dc4 NH |
1833 | return -EBUSY; |
1834 | } | |
98fd1ef4 | 1835 | unlock_page(hpage); |
acc14dc4 NH |
1836 | get_hwpoison_page(page); |
1837 | put_hwpoison_page(hpage); | |
1838 | } | |
1839 | ||
d4ae9916 NH |
1840 | /* |
1841 | * Setting MIGRATE_ISOLATE here ensures that the page will be linked | |
1842 | * to free list immediately (not via pcplist) when released after | |
1843 | * successful page migration. Otherwise we can't guarantee that the | |
1844 | * page is really free after put_page() returns, so | |
1845 | * set_hwpoison_free_buddy_page() highly likely fails. | |
1846 | */ | |
1847 | mt = get_pageblock_migratetype(page); | |
1848 | set_pageblock_migratetype(page, MIGRATE_ISOLATE); | |
acc14dc4 NH |
1849 | if (PageHuge(page)) |
1850 | ret = soft_offline_huge_page(page, flags); | |
1851 | else | |
1852 | ret = __soft_offline_page(page, flags); | |
d4ae9916 | 1853 | set_pageblock_migratetype(page, mt); |
acc14dc4 NH |
1854 | return ret; |
1855 | } | |
1856 | ||
d4ae9916 | 1857 | static int soft_offline_free_page(struct page *page) |
acc14dc4 | 1858 | { |
6bc9b564 | 1859 | int rc = 0; |
b37ff71c | 1860 | struct page *head = compound_head(page); |
acc14dc4 | 1861 | |
6bc9b564 NH |
1862 | if (PageHuge(head)) |
1863 | rc = dissolve_free_huge_page(page); | |
d4ae9916 NH |
1864 | if (!rc) { |
1865 | if (set_hwpoison_free_buddy_page(page)) | |
1866 | num_poisoned_pages_inc(); | |
1867 | else | |
1868 | rc = -EBUSY; | |
1869 | } | |
1870 | return rc; | |
acc14dc4 NH |
1871 | } |
1872 | ||
86e05773 WL |
1873 | /** |
1874 | * soft_offline_page - Soft offline a page. | |
1875 | * @page: page to offline | |
1876 | * @flags: flags. Same as memory_failure(). | |
1877 | * | |
1878 | * Returns 0 on success, otherwise negated errno. | |
1879 | * | |
1880 | * Soft offline a page, by migration or invalidation, | |
1881 | * without killing anything. This is for the case when | |
1882 | * a page is not corrupted yet (so it's still valid to access), | |
1883 | * but has had a number of corrected errors and is better taken | |
1884 | * out. | |
1885 | * | |
1886 | * The actual policy on when to do that is maintained by | |
1887 | * user space. | |
1888 | * | |
1889 | * This should never impact any application or cause data loss, | |
1890 | * however it might take some time. | |
1891 | * | |
1892 | * This is not a 100% solution for all memory, but tries to be | |
1893 | * ``good enough'' for the majority of memory. | |
1894 | */ | |
1895 | int soft_offline_page(struct page *page, int flags) | |
1896 | { | |
1897 | int ret; | |
1898 | unsigned long pfn = page_to_pfn(page); | |
86e05773 | 1899 | |
86a66810 DW |
1900 | if (is_zone_device_page(page)) { |
1901 | pr_debug_ratelimited("soft_offline: %#lx page is device page\n", | |
1902 | pfn); | |
1903 | if (flags & MF_COUNT_INCREASED) | |
1904 | put_page(page); | |
1905 | return -EIO; | |
1906 | } | |
1907 | ||
86e05773 WL |
1908 | if (PageHWPoison(page)) { |
1909 | pr_info("soft offline: %#lx page already poisoned\n", pfn); | |
1e0e635b | 1910 | if (flags & MF_COUNT_INCREASED) |
665d9da7 | 1911 | put_hwpoison_page(page); |
86e05773 WL |
1912 | return -EBUSY; |
1913 | } | |
86e05773 | 1914 | |
bfc8c901 | 1915 | get_online_mems(); |
86e05773 | 1916 | ret = get_any_page(page, pfn, flags); |
bfc8c901 | 1917 | put_online_mems(); |
4e41a30c | 1918 | |
acc14dc4 NH |
1919 | if (ret > 0) |
1920 | ret = soft_offline_in_use_page(page, flags); | |
1921 | else if (ret == 0) | |
d4ae9916 | 1922 | ret = soft_offline_free_page(page); |
4e41a30c | 1923 | |
86e05773 WL |
1924 | return ret; |
1925 | } |