]>
Commit | Line | Data |
---|---|---|
1439f94c | 1 | // SPDX-License-Identifier: GPL-2.0-only |
6a46079c AK |
2 | /* |
3 | * Copyright (C) 2008, 2009 Intel Corporation | |
4 | * Authors: Andi Kleen, Fengguang Wu | |
5 | * | |
6a46079c | 6 | * High level machine check handler. Handles pages reported by the |
1c80b990 | 7 | * hardware as being corrupted usually due to a multi-bit ECC memory or cache |
6a46079c | 8 | * failure. |
1c80b990 AK |
9 | * |
10 | * In addition there is a "soft offline" entry point that allows stop using | |
11 | * not-yet-corrupted-by-suspicious pages without killing anything. | |
6a46079c AK |
12 | * |
13 | * Handles page cache pages in various states. The tricky part | |
1c80b990 AK |
14 | * here is that we can access any page asynchronously in respect to |
15 | * other VM users, because memory failures could happen anytime and | |
16 | * anywhere. This could violate some of their assumptions. This is why | |
17 | * this code has to be extremely careful. Generally it tries to use | |
18 | * normal locking rules, as in get the standard locks, even if that means | |
19 | * the error handling takes potentially a long time. | |
e0de78df AK |
20 | * |
21 | * It can be very tempting to add handling for obscure cases here. | |
22 | * In general any code for handling new cases should only be added iff: | |
23 | * - You know how to test it. | |
24 | * - You have a test that can be added to mce-test | |
25 | * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/ | |
26 | * - The case actually shows up as a frequent (top 10) page state in | |
27 | * tools/vm/page-types when running a real workload. | |
1c80b990 AK |
28 | * |
29 | * There are several operations here with exponential complexity because | |
30 | * of unsuitable VM data structures. For example the operation to map back | |
31 | * from RMAP chains to processes has to walk the complete process list and | |
32 | * has non linear complexity with the number. But since memory corruptions | |
33 | * are rare we hope to get away with this. This avoids impacting the core | |
34 | * VM. | |
6a46079c | 35 | */ |
6a46079c AK |
36 | #include <linux/kernel.h> |
37 | #include <linux/mm.h> | |
38 | #include <linux/page-flags.h> | |
478c5ffc | 39 | #include <linux/kernel-page-flags.h> |
3f07c014 | 40 | #include <linux/sched/signal.h> |
29930025 | 41 | #include <linux/sched/task.h> |
01e00f88 | 42 | #include <linux/ksm.h> |
6a46079c | 43 | #include <linux/rmap.h> |
b9e15baf | 44 | #include <linux/export.h> |
6a46079c AK |
45 | #include <linux/pagemap.h> |
46 | #include <linux/swap.h> | |
47 | #include <linux/backing-dev.h> | |
facb6011 | 48 | #include <linux/migrate.h> |
facb6011 | 49 | #include <linux/suspend.h> |
5a0e3ad6 | 50 | #include <linux/slab.h> |
bf998156 | 51 | #include <linux/swapops.h> |
7af446a8 | 52 | #include <linux/hugetlb.h> |
20d6c96b | 53 | #include <linux/memory_hotplug.h> |
5db8a73a | 54 | #include <linux/mm_inline.h> |
6100e34b | 55 | #include <linux/memremap.h> |
ea8f5fb8 | 56 | #include <linux/kfifo.h> |
a5f65109 | 57 | #include <linux/ratelimit.h> |
d4ae9916 | 58 | #include <linux/page-isolation.h> |
6a46079c | 59 | #include "internal.h" |
97f0b134 | 60 | #include "ras/ras_event.h" |
6a46079c AK |
61 | |
62 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
63 | ||
64 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
65 | ||
293c07e3 | 66 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 67 | |
06be6ff3 OS |
68 | static void page_handle_poison(struct page *page) |
69 | { | |
70 | SetPageHWPoison(page); | |
71 | page_ref_inc(page); | |
72 | num_poisoned_pages_inc(); | |
73 | } | |
74 | ||
27df5068 AK |
75 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
76 | ||
1bfe5feb | 77 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
78 | u32 hwpoison_filter_dev_major = ~0U; |
79 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
80 | u64 hwpoison_filter_flags_mask; |
81 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 82 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
83 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
84 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
85 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
86 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
87 | |
88 | static int hwpoison_filter_dev(struct page *p) | |
89 | { | |
90 | struct address_space *mapping; | |
91 | dev_t dev; | |
92 | ||
93 | if (hwpoison_filter_dev_major == ~0U && | |
94 | hwpoison_filter_dev_minor == ~0U) | |
95 | return 0; | |
96 | ||
97 | /* | |
1c80b990 | 98 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
99 | */ |
100 | if (PageSlab(p)) | |
101 | return -EINVAL; | |
102 | ||
103 | mapping = page_mapping(p); | |
104 | if (mapping == NULL || mapping->host == NULL) | |
105 | return -EINVAL; | |
106 | ||
107 | dev = mapping->host->i_sb->s_dev; | |
108 | if (hwpoison_filter_dev_major != ~0U && | |
109 | hwpoison_filter_dev_major != MAJOR(dev)) | |
110 | return -EINVAL; | |
111 | if (hwpoison_filter_dev_minor != ~0U && | |
112 | hwpoison_filter_dev_minor != MINOR(dev)) | |
113 | return -EINVAL; | |
114 | ||
115 | return 0; | |
116 | } | |
117 | ||
478c5ffc WF |
118 | static int hwpoison_filter_flags(struct page *p) |
119 | { | |
120 | if (!hwpoison_filter_flags_mask) | |
121 | return 0; | |
122 | ||
123 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
124 | hwpoison_filter_flags_value) | |
125 | return 0; | |
126 | else | |
127 | return -EINVAL; | |
128 | } | |
129 | ||
4fd466eb AK |
130 | /* |
131 | * This allows stress tests to limit test scope to a collection of tasks | |
132 | * by putting them under some memcg. This prevents killing unrelated/important | |
133 | * processes such as /sbin/init. Note that the target task may share clean | |
134 | * pages with init (eg. libc text), which is harmless. If the target task | |
135 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
136 | * is also included in the memcg. At last, due to race conditions this filter | |
137 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
138 | * a freed page. | |
139 | */ | |
94a59fb3 | 140 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
141 | u64 hwpoison_filter_memcg; |
142 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
143 | static int hwpoison_filter_task(struct page *p) | |
144 | { | |
4fd466eb AK |
145 | if (!hwpoison_filter_memcg) |
146 | return 0; | |
147 | ||
94a59fb3 | 148 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
149 | return -EINVAL; |
150 | ||
151 | return 0; | |
152 | } | |
153 | #else | |
154 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
155 | #endif | |
156 | ||
7c116f2b WF |
157 | int hwpoison_filter(struct page *p) |
158 | { | |
1bfe5feb HL |
159 | if (!hwpoison_filter_enable) |
160 | return 0; | |
161 | ||
7c116f2b WF |
162 | if (hwpoison_filter_dev(p)) |
163 | return -EINVAL; | |
164 | ||
478c5ffc WF |
165 | if (hwpoison_filter_flags(p)) |
166 | return -EINVAL; | |
167 | ||
4fd466eb AK |
168 | if (hwpoison_filter_task(p)) |
169 | return -EINVAL; | |
170 | ||
7c116f2b WF |
171 | return 0; |
172 | } | |
27df5068 AK |
173 | #else |
174 | int hwpoison_filter(struct page *p) | |
175 | { | |
176 | return 0; | |
177 | } | |
178 | #endif | |
179 | ||
7c116f2b WF |
180 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
181 | ||
ae1139ec DW |
182 | /* |
183 | * Kill all processes that have a poisoned page mapped and then isolate | |
184 | * the page. | |
185 | * | |
186 | * General strategy: | |
187 | * Find all processes having the page mapped and kill them. | |
188 | * But we keep a page reference around so that the page is not | |
189 | * actually freed yet. | |
190 | * Then stash the page away | |
191 | * | |
192 | * There's no convenient way to get back to mapped processes | |
193 | * from the VMAs. So do a brute-force search over all | |
194 | * running processes. | |
195 | * | |
196 | * Remember that machine checks are not common (or rather | |
197 | * if they are common you have other problems), so this shouldn't | |
198 | * be a performance issue. | |
199 | * | |
200 | * Also there are some races possible while we get from the | |
201 | * error detection to actually handle it. | |
202 | */ | |
203 | ||
204 | struct to_kill { | |
205 | struct list_head nd; | |
206 | struct task_struct *tsk; | |
207 | unsigned long addr; | |
208 | short size_shift; | |
ae1139ec DW |
209 | }; |
210 | ||
6a46079c | 211 | /* |
7329bbeb TL |
212 | * Send all the processes who have the page mapped a signal. |
213 | * ``action optional'' if they are not immediately affected by the error | |
214 | * ``action required'' if error happened in current execution context | |
6a46079c | 215 | */ |
ae1139ec | 216 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 217 | { |
ae1139ec DW |
218 | struct task_struct *t = tk->tsk; |
219 | short addr_lsb = tk->size_shift; | |
872e9a20 | 220 | int ret = 0; |
6a46079c | 221 | |
03151c6e | 222 | pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", |
872e9a20 | 223 | pfn, t->comm, t->pid); |
7329bbeb | 224 | |
872e9a20 | 225 | if (flags & MF_ACTION_REQUIRED) { |
03151c6e NH |
226 | WARN_ON_ONCE(t != current); |
227 | ret = force_sig_mceerr(BUS_MCEERR_AR, | |
872e9a20 | 228 | (void __user *)tk->addr, addr_lsb); |
7329bbeb TL |
229 | } else { |
230 | /* | |
231 | * Don't use force here, it's convenient if the signal | |
232 | * can be temporarily blocked. | |
233 | * This could cause a loop when the user sets SIGBUS | |
234 | * to SIG_IGN, but hopefully no one will do that? | |
235 | */ | |
ae1139ec | 236 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
c0f45555 | 237 | addr_lsb, t); /* synchronous? */ |
7329bbeb | 238 | } |
6a46079c | 239 | if (ret < 0) |
495367c0 | 240 | pr_info("Memory failure: Error sending signal to %s:%d: %d\n", |
1170532b | 241 | t->comm, t->pid, ret); |
6a46079c AK |
242 | return ret; |
243 | } | |
244 | ||
588f9ce6 AK |
245 | /* |
246 | * When a unknown page type is encountered drain as many buffers as possible | |
247 | * in the hope to turn the page into a LRU or free page, which we can handle. | |
248 | */ | |
facb6011 | 249 | void shake_page(struct page *p, int access) |
588f9ce6 | 250 | { |
8bcb74de NH |
251 | if (PageHuge(p)) |
252 | return; | |
253 | ||
588f9ce6 AK |
254 | if (!PageSlab(p)) { |
255 | lru_add_drain_all(); | |
256 | if (PageLRU(p)) | |
257 | return; | |
c0554329 | 258 | drain_all_pages(page_zone(p)); |
588f9ce6 AK |
259 | if (PageLRU(p) || is_free_buddy_page(p)) |
260 | return; | |
261 | } | |
facb6011 | 262 | |
588f9ce6 | 263 | /* |
6b4f7799 JW |
264 | * Only call shrink_node_slabs here (which would also shrink |
265 | * other caches) if access is not potentially fatal. | |
588f9ce6 | 266 | */ |
cb731d6c VD |
267 | if (access) |
268 | drop_slab_node(page_to_nid(p)); | |
588f9ce6 AK |
269 | } |
270 | EXPORT_SYMBOL_GPL(shake_page); | |
271 | ||
6100e34b DW |
272 | static unsigned long dev_pagemap_mapping_shift(struct page *page, |
273 | struct vm_area_struct *vma) | |
274 | { | |
275 | unsigned long address = vma_address(page, vma); | |
276 | pgd_t *pgd; | |
277 | p4d_t *p4d; | |
278 | pud_t *pud; | |
279 | pmd_t *pmd; | |
280 | pte_t *pte; | |
281 | ||
282 | pgd = pgd_offset(vma->vm_mm, address); | |
283 | if (!pgd_present(*pgd)) | |
284 | return 0; | |
285 | p4d = p4d_offset(pgd, address); | |
286 | if (!p4d_present(*p4d)) | |
287 | return 0; | |
288 | pud = pud_offset(p4d, address); | |
289 | if (!pud_present(*pud)) | |
290 | return 0; | |
291 | if (pud_devmap(*pud)) | |
292 | return PUD_SHIFT; | |
293 | pmd = pmd_offset(pud, address); | |
294 | if (!pmd_present(*pmd)) | |
295 | return 0; | |
296 | if (pmd_devmap(*pmd)) | |
297 | return PMD_SHIFT; | |
298 | pte = pte_offset_map(pmd, address); | |
299 | if (!pte_present(*pte)) | |
300 | return 0; | |
301 | if (pte_devmap(*pte)) | |
302 | return PAGE_SHIFT; | |
303 | return 0; | |
304 | } | |
6a46079c AK |
305 | |
306 | /* | |
307 | * Failure handling: if we can't find or can't kill a process there's | |
308 | * not much we can do. We just print a message and ignore otherwise. | |
309 | */ | |
310 | ||
311 | /* | |
312 | * Schedule a process for later kill. | |
313 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
6a46079c AK |
314 | */ |
315 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
316 | struct vm_area_struct *vma, | |
996ff7a0 | 317 | struct list_head *to_kill) |
6a46079c AK |
318 | { |
319 | struct to_kill *tk; | |
320 | ||
996ff7a0 JC |
321 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
322 | if (!tk) { | |
323 | pr_err("Memory failure: Out of memory while machine check handling\n"); | |
324 | return; | |
6a46079c | 325 | } |
996ff7a0 | 326 | |
6a46079c | 327 | tk->addr = page_address_in_vma(p, vma); |
6100e34b DW |
328 | if (is_zone_device_page(p)) |
329 | tk->size_shift = dev_pagemap_mapping_shift(p, vma); | |
330 | else | |
75068518 | 331 | tk->size_shift = page_shift(compound_head(p)); |
6a46079c AK |
332 | |
333 | /* | |
3d7fed4a JC |
334 | * Send SIGKILL if "tk->addr == -EFAULT". Also, as |
335 | * "tk->size_shift" is always non-zero for !is_zone_device_page(), | |
336 | * so "tk->size_shift == 0" effectively checks no mapping on | |
337 | * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times | |
338 | * to a process' address space, it's possible not all N VMAs | |
339 | * contain mappings for the page, but at least one VMA does. | |
340 | * Only deliver SIGBUS with payload derived from the VMA that | |
341 | * has a mapping for the page. | |
6a46079c | 342 | */ |
3d7fed4a | 343 | if (tk->addr == -EFAULT) { |
495367c0 | 344 | pr_info("Memory failure: Unable to find user space address %lx in %s\n", |
6a46079c | 345 | page_to_pfn(p), tsk->comm); |
3d7fed4a JC |
346 | } else if (tk->size_shift == 0) { |
347 | kfree(tk); | |
348 | return; | |
6a46079c | 349 | } |
996ff7a0 | 350 | |
6a46079c AK |
351 | get_task_struct(tsk); |
352 | tk->tsk = tsk; | |
353 | list_add_tail(&tk->nd, to_kill); | |
354 | } | |
355 | ||
356 | /* | |
357 | * Kill the processes that have been collected earlier. | |
358 | * | |
359 | * Only do anything when DOIT is set, otherwise just free the list | |
360 | * (this is used for clean pages which do not need killing) | |
361 | * Also when FAIL is set do a force kill because something went | |
362 | * wrong earlier. | |
363 | */ | |
ae1139ec DW |
364 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
365 | unsigned long pfn, int flags) | |
6a46079c AK |
366 | { |
367 | struct to_kill *tk, *next; | |
368 | ||
369 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 370 | if (forcekill) { |
6a46079c | 371 | /* |
af901ca1 | 372 | * In case something went wrong with munmapping |
6a46079c AK |
373 | * make sure the process doesn't catch the |
374 | * signal and then access the memory. Just kill it. | |
6a46079c | 375 | */ |
3d7fed4a | 376 | if (fail || tk->addr == -EFAULT) { |
495367c0 | 377 | pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 378 | pfn, tk->tsk->comm, tk->tsk->pid); |
6376360e NH |
379 | do_send_sig_info(SIGKILL, SEND_SIG_PRIV, |
380 | tk->tsk, PIDTYPE_PID); | |
6a46079c AK |
381 | } |
382 | ||
383 | /* | |
384 | * In theory the process could have mapped | |
385 | * something else on the address in-between. We could | |
386 | * check for that, but we need to tell the | |
387 | * process anyways. | |
388 | */ | |
ae1139ec | 389 | else if (kill_proc(tk, pfn, flags) < 0) |
495367c0 | 390 | pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 391 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c AK |
392 | } |
393 | put_task_struct(tk->tsk); | |
394 | kfree(tk); | |
395 | } | |
396 | } | |
397 | ||
3ba08129 NH |
398 | /* |
399 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
400 | * on behalf of the thread group. Return task_struct of the (first found) | |
401 | * dedicated thread if found, and return NULL otherwise. | |
402 | * | |
403 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
404 | * have to call rcu_read_lock/unlock() in this function. | |
405 | */ | |
406 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 407 | { |
3ba08129 NH |
408 | struct task_struct *t; |
409 | ||
4e018b45 NH |
410 | for_each_thread(tsk, t) { |
411 | if (t->flags & PF_MCE_PROCESS) { | |
412 | if (t->flags & PF_MCE_EARLY) | |
413 | return t; | |
414 | } else { | |
415 | if (sysctl_memory_failure_early_kill) | |
416 | return t; | |
417 | } | |
418 | } | |
3ba08129 NH |
419 | return NULL; |
420 | } | |
421 | ||
422 | /* | |
423 | * Determine whether a given process is "early kill" process which expects | |
424 | * to be signaled when some page under the process is hwpoisoned. | |
425 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
426 | * specified) if the process is "early kill," and otherwise returns NULL. | |
03151c6e NH |
427 | * |
428 | * Note that the above is true for Action Optional case, but not for Action | |
429 | * Required case where SIGBUS should sent only to the current thread. | |
3ba08129 NH |
430 | */ |
431 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
432 | int force_early) | |
433 | { | |
6a46079c | 434 | if (!tsk->mm) |
3ba08129 | 435 | return NULL; |
03151c6e NH |
436 | if (force_early) { |
437 | /* | |
438 | * Comparing ->mm here because current task might represent | |
439 | * a subthread, while tsk always points to the main thread. | |
440 | */ | |
441 | if (tsk->mm == current->mm) | |
442 | return current; | |
443 | else | |
444 | return NULL; | |
445 | } | |
4e018b45 | 446 | return find_early_kill_thread(tsk); |
6a46079c AK |
447 | } |
448 | ||
449 | /* | |
450 | * Collect processes when the error hit an anonymous page. | |
451 | */ | |
452 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 453 | int force_early) |
6a46079c AK |
454 | { |
455 | struct vm_area_struct *vma; | |
456 | struct task_struct *tsk; | |
457 | struct anon_vma *av; | |
bf181b9f | 458 | pgoff_t pgoff; |
6a46079c | 459 | |
4fc3f1d6 | 460 | av = page_lock_anon_vma_read(page); |
6a46079c | 461 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
462 | return; |
463 | ||
a0f7a756 | 464 | pgoff = page_to_pgoff(page); |
9b679320 | 465 | read_lock(&tasklist_lock); |
6a46079c | 466 | for_each_process (tsk) { |
5beb4930 | 467 | struct anon_vma_chain *vmac; |
3ba08129 | 468 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 469 | |
3ba08129 | 470 | if (!t) |
6a46079c | 471 | continue; |
bf181b9f ML |
472 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
473 | pgoff, pgoff) { | |
5beb4930 | 474 | vma = vmac->vma; |
6a46079c AK |
475 | if (!page_mapped_in_vma(page, vma)) |
476 | continue; | |
3ba08129 | 477 | if (vma->vm_mm == t->mm) |
996ff7a0 | 478 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
479 | } |
480 | } | |
6a46079c | 481 | read_unlock(&tasklist_lock); |
4fc3f1d6 | 482 | page_unlock_anon_vma_read(av); |
6a46079c AK |
483 | } |
484 | ||
485 | /* | |
486 | * Collect processes when the error hit a file mapped page. | |
487 | */ | |
488 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 489 | int force_early) |
6a46079c AK |
490 | { |
491 | struct vm_area_struct *vma; | |
492 | struct task_struct *tsk; | |
6a46079c | 493 | struct address_space *mapping = page->mapping; |
c43bc03d | 494 | pgoff_t pgoff; |
6a46079c | 495 | |
d28eb9c8 | 496 | i_mmap_lock_read(mapping); |
9b679320 | 497 | read_lock(&tasklist_lock); |
c43bc03d | 498 | pgoff = page_to_pgoff(page); |
6a46079c | 499 | for_each_process(tsk) { |
3ba08129 | 500 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 501 | |
3ba08129 | 502 | if (!t) |
6a46079c | 503 | continue; |
6b2dbba8 | 504 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
505 | pgoff) { |
506 | /* | |
507 | * Send early kill signal to tasks where a vma covers | |
508 | * the page but the corrupted page is not necessarily | |
509 | * mapped it in its pte. | |
510 | * Assume applications who requested early kill want | |
511 | * to be informed of all such data corruptions. | |
512 | */ | |
3ba08129 | 513 | if (vma->vm_mm == t->mm) |
996ff7a0 | 514 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
515 | } |
516 | } | |
6a46079c | 517 | read_unlock(&tasklist_lock); |
d28eb9c8 | 518 | i_mmap_unlock_read(mapping); |
6a46079c AK |
519 | } |
520 | ||
521 | /* | |
522 | * Collect the processes who have the corrupted page mapped to kill. | |
6a46079c | 523 | */ |
74614de1 TL |
524 | static void collect_procs(struct page *page, struct list_head *tokill, |
525 | int force_early) | |
6a46079c | 526 | { |
6a46079c AK |
527 | if (!page->mapping) |
528 | return; | |
529 | ||
6a46079c | 530 | if (PageAnon(page)) |
996ff7a0 | 531 | collect_procs_anon(page, tokill, force_early); |
6a46079c | 532 | else |
996ff7a0 | 533 | collect_procs_file(page, tokill, force_early); |
6a46079c AK |
534 | } |
535 | ||
6a46079c | 536 | static const char *action_name[] = { |
cc637b17 XX |
537 | [MF_IGNORED] = "Ignored", |
538 | [MF_FAILED] = "Failed", | |
539 | [MF_DELAYED] = "Delayed", | |
540 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
541 | }; |
542 | ||
543 | static const char * const action_page_types[] = { | |
cc637b17 XX |
544 | [MF_MSG_KERNEL] = "reserved kernel page", |
545 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
546 | [MF_MSG_SLAB] = "kernel slab page", | |
547 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
548 | [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", | |
549 | [MF_MSG_HUGE] = "huge page", | |
550 | [MF_MSG_FREE_HUGE] = "free huge page", | |
31286a84 | 551 | [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page", |
cc637b17 XX |
552 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", |
553 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
554 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
555 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
556 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
557 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
558 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
559 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
560 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
561 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
562 | [MF_MSG_BUDDY] = "free buddy page", | |
563 | [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", | |
6100e34b | 564 | [MF_MSG_DAX] = "dax page", |
cc637b17 | 565 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
566 | }; |
567 | ||
dc2a1cbf WF |
568 | /* |
569 | * XXX: It is possible that a page is isolated from LRU cache, | |
570 | * and then kept in swap cache or failed to remove from page cache. | |
571 | * The page count will stop it from being freed by unpoison. | |
572 | * Stress tests should be aware of this memory leak problem. | |
573 | */ | |
574 | static int delete_from_lru_cache(struct page *p) | |
575 | { | |
576 | if (!isolate_lru_page(p)) { | |
577 | /* | |
578 | * Clear sensible page flags, so that the buddy system won't | |
579 | * complain when the page is unpoison-and-freed. | |
580 | */ | |
581 | ClearPageActive(p); | |
582 | ClearPageUnevictable(p); | |
18365225 MH |
583 | |
584 | /* | |
585 | * Poisoned page might never drop its ref count to 0 so we have | |
586 | * to uncharge it manually from its memcg. | |
587 | */ | |
588 | mem_cgroup_uncharge(p); | |
589 | ||
dc2a1cbf WF |
590 | /* |
591 | * drop the page count elevated by isolate_lru_page() | |
592 | */ | |
09cbfeaf | 593 | put_page(p); |
dc2a1cbf WF |
594 | return 0; |
595 | } | |
596 | return -EIO; | |
597 | } | |
598 | ||
78bb9203 NH |
599 | static int truncate_error_page(struct page *p, unsigned long pfn, |
600 | struct address_space *mapping) | |
601 | { | |
602 | int ret = MF_FAILED; | |
603 | ||
604 | if (mapping->a_ops->error_remove_page) { | |
605 | int err = mapping->a_ops->error_remove_page(mapping, p); | |
606 | ||
607 | if (err != 0) { | |
608 | pr_info("Memory failure: %#lx: Failed to punch page: %d\n", | |
609 | pfn, err); | |
610 | } else if (page_has_private(p) && | |
611 | !try_to_release_page(p, GFP_NOIO)) { | |
612 | pr_info("Memory failure: %#lx: failed to release buffers\n", | |
613 | pfn); | |
614 | } else { | |
615 | ret = MF_RECOVERED; | |
616 | } | |
617 | } else { | |
618 | /* | |
619 | * If the file system doesn't support it just invalidate | |
620 | * This fails on dirty or anything with private pages | |
621 | */ | |
622 | if (invalidate_inode_page(p)) | |
623 | ret = MF_RECOVERED; | |
624 | else | |
625 | pr_info("Memory failure: %#lx: Failed to invalidate\n", | |
626 | pfn); | |
627 | } | |
628 | ||
629 | return ret; | |
630 | } | |
631 | ||
6a46079c AK |
632 | /* |
633 | * Error hit kernel page. | |
634 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
635 | * could be more sophisticated. | |
636 | */ | |
637 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c | 638 | { |
cc637b17 | 639 | return MF_IGNORED; |
6a46079c AK |
640 | } |
641 | ||
642 | /* | |
643 | * Page in unknown state. Do nothing. | |
644 | */ | |
645 | static int me_unknown(struct page *p, unsigned long pfn) | |
646 | { | |
495367c0 | 647 | pr_err("Memory failure: %#lx: Unknown page state\n", pfn); |
cc637b17 | 648 | return MF_FAILED; |
6a46079c AK |
649 | } |
650 | ||
6a46079c AK |
651 | /* |
652 | * Clean (or cleaned) page cache page. | |
653 | */ | |
654 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
655 | { | |
6a46079c AK |
656 | struct address_space *mapping; |
657 | ||
dc2a1cbf WF |
658 | delete_from_lru_cache(p); |
659 | ||
6a46079c AK |
660 | /* |
661 | * For anonymous pages we're done the only reference left | |
662 | * should be the one m_f() holds. | |
663 | */ | |
664 | if (PageAnon(p)) | |
cc637b17 | 665 | return MF_RECOVERED; |
6a46079c AK |
666 | |
667 | /* | |
668 | * Now truncate the page in the page cache. This is really | |
669 | * more like a "temporary hole punch" | |
670 | * Don't do this for block devices when someone else | |
671 | * has a reference, because it could be file system metadata | |
672 | * and that's not safe to truncate. | |
673 | */ | |
674 | mapping = page_mapping(p); | |
675 | if (!mapping) { | |
676 | /* | |
677 | * Page has been teared down in the meanwhile | |
678 | */ | |
cc637b17 | 679 | return MF_FAILED; |
6a46079c AK |
680 | } |
681 | ||
682 | /* | |
683 | * Truncation is a bit tricky. Enable it per file system for now. | |
684 | * | |
685 | * Open: to take i_mutex or not for this? Right now we don't. | |
686 | */ | |
78bb9203 | 687 | return truncate_error_page(p, pfn, mapping); |
6a46079c AK |
688 | } |
689 | ||
690 | /* | |
549543df | 691 | * Dirty pagecache page |
6a46079c AK |
692 | * Issues: when the error hit a hole page the error is not properly |
693 | * propagated. | |
694 | */ | |
695 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
696 | { | |
697 | struct address_space *mapping = page_mapping(p); | |
698 | ||
699 | SetPageError(p); | |
700 | /* TBD: print more information about the file. */ | |
701 | if (mapping) { | |
702 | /* | |
703 | * IO error will be reported by write(), fsync(), etc. | |
704 | * who check the mapping. | |
705 | * This way the application knows that something went | |
706 | * wrong with its dirty file data. | |
707 | * | |
708 | * There's one open issue: | |
709 | * | |
710 | * The EIO will be only reported on the next IO | |
711 | * operation and then cleared through the IO map. | |
712 | * Normally Linux has two mechanisms to pass IO error | |
713 | * first through the AS_EIO flag in the address space | |
714 | * and then through the PageError flag in the page. | |
715 | * Since we drop pages on memory failure handling the | |
716 | * only mechanism open to use is through AS_AIO. | |
717 | * | |
718 | * This has the disadvantage that it gets cleared on | |
719 | * the first operation that returns an error, while | |
720 | * the PageError bit is more sticky and only cleared | |
721 | * when the page is reread or dropped. If an | |
722 | * application assumes it will always get error on | |
723 | * fsync, but does other operations on the fd before | |
25985edc | 724 | * and the page is dropped between then the error |
6a46079c AK |
725 | * will not be properly reported. |
726 | * | |
727 | * This can already happen even without hwpoisoned | |
728 | * pages: first on metadata IO errors (which only | |
729 | * report through AS_EIO) or when the page is dropped | |
730 | * at the wrong time. | |
731 | * | |
732 | * So right now we assume that the application DTRT on | |
733 | * the first EIO, but we're not worse than other parts | |
734 | * of the kernel. | |
735 | */ | |
af21bfaf | 736 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
737 | } |
738 | ||
739 | return me_pagecache_clean(p, pfn); | |
740 | } | |
741 | ||
742 | /* | |
743 | * Clean and dirty swap cache. | |
744 | * | |
745 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
746 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
747 | * referenced concurrently by 2 types of PTEs: | |
748 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
749 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
750 | * and then | |
751 | * - clear dirty bit to prevent IO | |
752 | * - remove from LRU | |
753 | * - but keep in the swap cache, so that when we return to it on | |
754 | * a later page fault, we know the application is accessing | |
755 | * corrupted data and shall be killed (we installed simple | |
756 | * interception code in do_swap_page to catch it). | |
757 | * | |
758 | * Clean swap cache pages can be directly isolated. A later page fault will | |
759 | * bring in the known good data from disk. | |
760 | */ | |
761 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
762 | { | |
6a46079c AK |
763 | ClearPageDirty(p); |
764 | /* Trigger EIO in shmem: */ | |
765 | ClearPageUptodate(p); | |
766 | ||
dc2a1cbf | 767 | if (!delete_from_lru_cache(p)) |
cc637b17 | 768 | return MF_DELAYED; |
dc2a1cbf | 769 | else |
cc637b17 | 770 | return MF_FAILED; |
6a46079c AK |
771 | } |
772 | ||
773 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
774 | { | |
6a46079c | 775 | delete_from_swap_cache(p); |
e43c3afb | 776 | |
dc2a1cbf | 777 | if (!delete_from_lru_cache(p)) |
cc637b17 | 778 | return MF_RECOVERED; |
dc2a1cbf | 779 | else |
cc637b17 | 780 | return MF_FAILED; |
6a46079c AK |
781 | } |
782 | ||
783 | /* | |
784 | * Huge pages. Needs work. | |
785 | * Issues: | |
93f70f90 NH |
786 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
787 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
788 | */ |
789 | static int me_huge_page(struct page *p, unsigned long pfn) | |
790 | { | |
6de2b1aa | 791 | int res = 0; |
93f70f90 | 792 | struct page *hpage = compound_head(p); |
78bb9203 | 793 | struct address_space *mapping; |
2491ffee NH |
794 | |
795 | if (!PageHuge(hpage)) | |
796 | return MF_DELAYED; | |
797 | ||
78bb9203 NH |
798 | mapping = page_mapping(hpage); |
799 | if (mapping) { | |
800 | res = truncate_error_page(hpage, pfn, mapping); | |
801 | } else { | |
802 | unlock_page(hpage); | |
803 | /* | |
804 | * migration entry prevents later access on error anonymous | |
805 | * hugepage, so we can free and dissolve it into buddy to | |
806 | * save healthy subpages. | |
807 | */ | |
808 | if (PageAnon(hpage)) | |
809 | put_page(hpage); | |
810 | dissolve_free_huge_page(p); | |
811 | res = MF_RECOVERED; | |
812 | lock_page(hpage); | |
93f70f90 | 813 | } |
78bb9203 NH |
814 | |
815 | return res; | |
6a46079c AK |
816 | } |
817 | ||
818 | /* | |
819 | * Various page states we can handle. | |
820 | * | |
821 | * A page state is defined by its current page->flags bits. | |
822 | * The table matches them in order and calls the right handler. | |
823 | * | |
824 | * This is quite tricky because we can access page at any time | |
25985edc | 825 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
826 | * |
827 | * This is not complete. More states could be added. | |
828 | * For any missing state don't attempt recovery. | |
829 | */ | |
830 | ||
831 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 832 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
833 | #define unevict (1UL << PG_unevictable) |
834 | #define mlock (1UL << PG_mlocked) | |
6a46079c | 835 | #define lru (1UL << PG_lru) |
6a46079c | 836 | #define head (1UL << PG_head) |
6a46079c | 837 | #define slab (1UL << PG_slab) |
6a46079c AK |
838 | #define reserved (1UL << PG_reserved) |
839 | ||
840 | static struct page_state { | |
841 | unsigned long mask; | |
842 | unsigned long res; | |
cc637b17 | 843 | enum mf_action_page_type type; |
6a46079c AK |
844 | int (*action)(struct page *p, unsigned long pfn); |
845 | } error_states[] = { | |
cc637b17 | 846 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
847 | /* |
848 | * free pages are specially detected outside this table: | |
849 | * PG_buddy pages only make a small fraction of all free pages. | |
850 | */ | |
6a46079c AK |
851 | |
852 | /* | |
853 | * Could in theory check if slab page is free or if we can drop | |
854 | * currently unused objects without touching them. But just | |
855 | * treat it as standard kernel for now. | |
856 | */ | |
cc637b17 | 857 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 858 | |
cc637b17 | 859 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 860 | |
cc637b17 XX |
861 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
862 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 863 | |
cc637b17 XX |
864 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
865 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 866 | |
cc637b17 XX |
867 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
868 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 869 | |
cc637b17 XX |
870 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
871 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
872 | |
873 | /* | |
874 | * Catchall entry: must be at end. | |
875 | */ | |
cc637b17 | 876 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
877 | }; |
878 | ||
2326c467 AK |
879 | #undef dirty |
880 | #undef sc | |
881 | #undef unevict | |
882 | #undef mlock | |
2326c467 | 883 | #undef lru |
2326c467 | 884 | #undef head |
2326c467 AK |
885 | #undef slab |
886 | #undef reserved | |
887 | ||
ff604cf6 NH |
888 | /* |
889 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
890 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
891 | */ | |
cc3e2af4 XX |
892 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
893 | enum mf_result result) | |
6a46079c | 894 | { |
97f0b134 XX |
895 | trace_memory_failure_event(pfn, type, result); |
896 | ||
495367c0 | 897 | pr_err("Memory failure: %#lx: recovery action for %s: %s\n", |
64d37a2b | 898 | pfn, action_page_types[type], action_name[result]); |
6a46079c AK |
899 | } |
900 | ||
901 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 902 | unsigned long pfn) |
6a46079c AK |
903 | { |
904 | int result; | |
7456b040 | 905 | int count; |
6a46079c AK |
906 | |
907 | result = ps->action(p, pfn); | |
7456b040 | 908 | |
bd1ce5f9 | 909 | count = page_count(p) - 1; |
cc637b17 | 910 | if (ps->action == me_swapcache_dirty && result == MF_DELAYED) |
138ce286 | 911 | count--; |
78bb9203 | 912 | if (count > 0) { |
495367c0 | 913 | pr_err("Memory failure: %#lx: %s still referenced by %d users\n", |
64d37a2b | 914 | pfn, action_page_types[ps->type], count); |
cc637b17 | 915 | result = MF_FAILED; |
138ce286 | 916 | } |
64d37a2b | 917 | action_result(pfn, ps->type, result); |
6a46079c AK |
918 | |
919 | /* Could do more checks here if page looks ok */ | |
920 | /* | |
921 | * Could adjust zone counters here to correct for the missing page. | |
922 | */ | |
923 | ||
cc637b17 | 924 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
925 | } |
926 | ||
ead07f6a NH |
927 | /** |
928 | * get_hwpoison_page() - Get refcount for memory error handling: | |
929 | * @page: raw error page (hit by memory error) | |
930 | * | |
931 | * Return: return 0 if failed to grab the refcount, otherwise true (some | |
932 | * non-zero value.) | |
933 | */ | |
7e27f22c | 934 | static int get_hwpoison_page(struct page *page) |
ead07f6a NH |
935 | { |
936 | struct page *head = compound_head(page); | |
937 | ||
4e41a30c | 938 | if (!PageHuge(head) && PageTransHuge(head)) { |
98ed2b00 NH |
939 | /* |
940 | * Non anonymous thp exists only in allocation/free time. We | |
941 | * can't handle such a case correctly, so let's give it up. | |
942 | * This should be better than triggering BUG_ON when kernel | |
943 | * tries to touch the "partially handled" page. | |
944 | */ | |
945 | if (!PageAnon(head)) { | |
495367c0 | 946 | pr_err("Memory failure: %#lx: non anonymous thp\n", |
98ed2b00 NH |
947 | page_to_pfn(page)); |
948 | return 0; | |
949 | } | |
ead07f6a NH |
950 | } |
951 | ||
c2e7e00b KK |
952 | if (get_page_unless_zero(head)) { |
953 | if (head == compound_head(page)) | |
954 | return 1; | |
955 | ||
495367c0 CY |
956 | pr_info("Memory failure: %#lx cannot catch tail\n", |
957 | page_to_pfn(page)); | |
c2e7e00b KK |
958 | put_page(head); |
959 | } | |
960 | ||
961 | return 0; | |
ead07f6a | 962 | } |
ead07f6a | 963 | |
6a46079c AK |
964 | /* |
965 | * Do all that is necessary to remove user space mappings. Unmap | |
966 | * the pages and send SIGBUS to the processes if the data was dirty. | |
967 | */ | |
666e5a40 | 968 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
83b57531 | 969 | int flags, struct page **hpagep) |
6a46079c | 970 | { |
a128ca71 | 971 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS; |
6a46079c AK |
972 | struct address_space *mapping; |
973 | LIST_HEAD(tokill); | |
c0d0381a | 974 | bool unmap_success = true; |
6751ed65 | 975 | int kill = 1, forcekill; |
54b9dd14 | 976 | struct page *hpage = *hpagep; |
286c469a | 977 | bool mlocked = PageMlocked(hpage); |
6a46079c | 978 | |
93a9eb39 NH |
979 | /* |
980 | * Here we are interested only in user-mapped pages, so skip any | |
981 | * other types of pages. | |
982 | */ | |
983 | if (PageReserved(p) || PageSlab(p)) | |
666e5a40 | 984 | return true; |
93a9eb39 | 985 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 986 | return true; |
6a46079c | 987 | |
6a46079c AK |
988 | /* |
989 | * This check implies we don't kill processes if their pages | |
990 | * are in the swap cache early. Those are always late kills. | |
991 | */ | |
7af446a8 | 992 | if (!page_mapped(hpage)) |
666e5a40 | 993 | return true; |
1668bfd5 | 994 | |
52089b14 | 995 | if (PageKsm(p)) { |
495367c0 | 996 | pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn); |
666e5a40 | 997 | return false; |
52089b14 | 998 | } |
6a46079c AK |
999 | |
1000 | if (PageSwapCache(p)) { | |
495367c0 CY |
1001 | pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n", |
1002 | pfn); | |
6a46079c AK |
1003 | ttu |= TTU_IGNORE_HWPOISON; |
1004 | } | |
1005 | ||
1006 | /* | |
1007 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1008 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1009 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1010 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1011 | */ |
7af446a8 | 1012 | mapping = page_mapping(hpage); |
6751ed65 | 1013 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
f56753ac | 1014 | mapping_can_writeback(mapping)) { |
7af446a8 NH |
1015 | if (page_mkclean(hpage)) { |
1016 | SetPageDirty(hpage); | |
6a46079c AK |
1017 | } else { |
1018 | kill = 0; | |
1019 | ttu |= TTU_IGNORE_HWPOISON; | |
495367c0 | 1020 | pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1021 | pfn); |
1022 | } | |
1023 | } | |
1024 | ||
1025 | /* | |
1026 | * First collect all the processes that have the page | |
1027 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1028 | * because ttu takes the rmap data structures down. | |
1029 | * | |
1030 | * Error handling: We ignore errors here because | |
1031 | * there's nothing that can be done. | |
1032 | */ | |
1033 | if (kill) | |
415c64c1 | 1034 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1035 | |
c0d0381a MK |
1036 | if (!PageHuge(hpage)) { |
1037 | unmap_success = try_to_unmap(hpage, ttu); | |
1038 | } else { | |
1039 | /* | |
1040 | * For hugetlb pages, try_to_unmap could potentially call | |
1041 | * huge_pmd_unshare. Because of this, take semaphore in | |
1042 | * write mode here and set TTU_RMAP_LOCKED to indicate we | |
1043 | * have taken the lock at this higer level. | |
1044 | * | |
1045 | * Note that the call to hugetlb_page_mapping_lock_write | |
1046 | * is necessary even if mapping is already set. It handles | |
1047 | * ugliness of potentially having to drop page lock to obtain | |
1048 | * i_mmap_rwsem. | |
1049 | */ | |
1050 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1051 | ||
1052 | if (mapping) { | |
1053 | unmap_success = try_to_unmap(hpage, | |
1054 | ttu|TTU_RMAP_LOCKED); | |
1055 | i_mmap_unlock_write(mapping); | |
1056 | } else { | |
1057 | pr_info("Memory failure: %#lx: could not find mapping for mapped huge page\n", | |
1058 | pfn); | |
1059 | unmap_success = false; | |
1060 | } | |
1061 | } | |
666e5a40 | 1062 | if (!unmap_success) |
495367c0 | 1063 | pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1064 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1065 | |
286c469a NH |
1066 | /* |
1067 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1068 | * shake_page() again to ensure that it's flushed. | |
1069 | */ | |
1070 | if (mlocked) | |
1071 | shake_page(hpage, 0); | |
1072 | ||
6a46079c AK |
1073 | /* |
1074 | * Now that the dirty bit has been propagated to the | |
1075 | * struct page and all unmaps done we can decide if | |
1076 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1077 | * was dirty or the process is not restartable, |
1078 | * otherwise the tokill list is merely | |
6a46079c AK |
1079 | * freed. When there was a problem unmapping earlier |
1080 | * use a more force-full uncatchable kill to prevent | |
1081 | * any accesses to the poisoned memory. | |
1082 | */ | |
415c64c1 | 1083 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
ae1139ec | 1084 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1085 | |
666e5a40 | 1086 | return unmap_success; |
6a46079c AK |
1087 | } |
1088 | ||
0348d2eb NH |
1089 | static int identify_page_state(unsigned long pfn, struct page *p, |
1090 | unsigned long page_flags) | |
761ad8d7 NH |
1091 | { |
1092 | struct page_state *ps; | |
0348d2eb NH |
1093 | |
1094 | /* | |
1095 | * The first check uses the current page flags which may not have any | |
1096 | * relevant information. The second check with the saved page flags is | |
1097 | * carried out only if the first check can't determine the page status. | |
1098 | */ | |
1099 | for (ps = error_states;; ps++) | |
1100 | if ((p->flags & ps->mask) == ps->res) | |
1101 | break; | |
1102 | ||
1103 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1104 | ||
1105 | if (!ps->mask) | |
1106 | for (ps = error_states;; ps++) | |
1107 | if ((page_flags & ps->mask) == ps->res) | |
1108 | break; | |
1109 | return page_action(ps, p, pfn); | |
1110 | } | |
1111 | ||
694bf0b0 OS |
1112 | static int try_to_split_thp_page(struct page *page, const char *msg) |
1113 | { | |
1114 | lock_page(page); | |
1115 | if (!PageAnon(page) || unlikely(split_huge_page(page))) { | |
1116 | unsigned long pfn = page_to_pfn(page); | |
1117 | ||
1118 | unlock_page(page); | |
1119 | if (!PageAnon(page)) | |
1120 | pr_info("%s: %#lx: non anonymous thp\n", msg, pfn); | |
1121 | else | |
1122 | pr_info("%s: %#lx: thp split failed\n", msg, pfn); | |
1123 | put_page(page); | |
1124 | return -EBUSY; | |
1125 | } | |
1126 | unlock_page(page); | |
1127 | ||
1128 | return 0; | |
1129 | } | |
1130 | ||
83b57531 | 1131 | static int memory_failure_hugetlb(unsigned long pfn, int flags) |
0348d2eb | 1132 | { |
761ad8d7 NH |
1133 | struct page *p = pfn_to_page(pfn); |
1134 | struct page *head = compound_head(p); | |
1135 | int res; | |
1136 | unsigned long page_flags; | |
1137 | ||
1138 | if (TestSetPageHWPoison(head)) { | |
1139 | pr_err("Memory failure: %#lx: already hardware poisoned\n", | |
1140 | pfn); | |
1141 | return 0; | |
1142 | } | |
1143 | ||
1144 | num_poisoned_pages_inc(); | |
1145 | ||
1146 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { | |
1147 | /* | |
1148 | * Check "filter hit" and "race with other subpage." | |
1149 | */ | |
1150 | lock_page(head); | |
1151 | if (PageHWPoison(head)) { | |
1152 | if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1153 | || (p != head && TestSetPageHWPoison(head))) { | |
1154 | num_poisoned_pages_dec(); | |
1155 | unlock_page(head); | |
1156 | return 0; | |
1157 | } | |
1158 | } | |
1159 | unlock_page(head); | |
1160 | dissolve_free_huge_page(p); | |
1161 | action_result(pfn, MF_MSG_FREE_HUGE, MF_DELAYED); | |
1162 | return 0; | |
1163 | } | |
1164 | ||
1165 | lock_page(head); | |
1166 | page_flags = head->flags; | |
1167 | ||
1168 | if (!PageHWPoison(head)) { | |
1169 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); | |
1170 | num_poisoned_pages_dec(); | |
1171 | unlock_page(head); | |
dd6e2402 | 1172 | put_page(head); |
761ad8d7 NH |
1173 | return 0; |
1174 | } | |
1175 | ||
31286a84 NH |
1176 | /* |
1177 | * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so | |
1178 | * simply disable it. In order to make it work properly, we need | |
1179 | * make sure that: | |
1180 | * - conversion of a pud that maps an error hugetlb into hwpoison | |
1181 | * entry properly works, and | |
1182 | * - other mm code walking over page table is aware of pud-aligned | |
1183 | * hwpoison entries. | |
1184 | */ | |
1185 | if (huge_page_size(page_hstate(head)) > PMD_SIZE) { | |
1186 | action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED); | |
1187 | res = -EBUSY; | |
1188 | goto out; | |
1189 | } | |
1190 | ||
83b57531 | 1191 | if (!hwpoison_user_mappings(p, pfn, flags, &head)) { |
761ad8d7 NH |
1192 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1193 | res = -EBUSY; | |
1194 | goto out; | |
1195 | } | |
1196 | ||
0348d2eb | 1197 | res = identify_page_state(pfn, p, page_flags); |
761ad8d7 NH |
1198 | out: |
1199 | unlock_page(head); | |
1200 | return res; | |
1201 | } | |
1202 | ||
6100e34b DW |
1203 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
1204 | struct dev_pagemap *pgmap) | |
1205 | { | |
1206 | struct page *page = pfn_to_page(pfn); | |
1207 | const bool unmap_success = true; | |
1208 | unsigned long size = 0; | |
1209 | struct to_kill *tk; | |
1210 | LIST_HEAD(tokill); | |
1211 | int rc = -EBUSY; | |
1212 | loff_t start; | |
27359fd6 | 1213 | dax_entry_t cookie; |
6100e34b DW |
1214 | |
1215 | /* | |
1216 | * Prevent the inode from being freed while we are interrogating | |
1217 | * the address_space, typically this would be handled by | |
1218 | * lock_page(), but dax pages do not use the page lock. This | |
1219 | * also prevents changes to the mapping of this pfn until | |
1220 | * poison signaling is complete. | |
1221 | */ | |
27359fd6 MW |
1222 | cookie = dax_lock_page(page); |
1223 | if (!cookie) | |
6100e34b DW |
1224 | goto out; |
1225 | ||
1226 | if (hwpoison_filter(page)) { | |
1227 | rc = 0; | |
1228 | goto unlock; | |
1229 | } | |
1230 | ||
25b2995a | 1231 | if (pgmap->type == MEMORY_DEVICE_PRIVATE) { |
6100e34b DW |
1232 | /* |
1233 | * TODO: Handle HMM pages which may need coordination | |
1234 | * with device-side memory. | |
1235 | */ | |
1236 | goto unlock; | |
6100e34b DW |
1237 | } |
1238 | ||
1239 | /* | |
1240 | * Use this flag as an indication that the dax page has been | |
1241 | * remapped UC to prevent speculative consumption of poison. | |
1242 | */ | |
1243 | SetPageHWPoison(page); | |
1244 | ||
1245 | /* | |
1246 | * Unlike System-RAM there is no possibility to swap in a | |
1247 | * different physical page at a given virtual address, so all | |
1248 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1249 | * SIGBUS (i.e. MF_MUST_KILL) | |
1250 | */ | |
1251 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1252 | collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED); | |
1253 | ||
1254 | list_for_each_entry(tk, &tokill, nd) | |
1255 | if (tk->size_shift) | |
1256 | size = max(size, 1UL << tk->size_shift); | |
1257 | if (size) { | |
1258 | /* | |
1259 | * Unmap the largest mapping to avoid breaking up | |
1260 | * device-dax mappings which are constant size. The | |
1261 | * actual size of the mapping being torn down is | |
1262 | * communicated in siginfo, see kill_proc() | |
1263 | */ | |
1264 | start = (page->index << PAGE_SHIFT) & ~(size - 1); | |
1265 | unmap_mapping_range(page->mapping, start, start + size, 0); | |
1266 | } | |
1267 | kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags); | |
1268 | rc = 0; | |
1269 | unlock: | |
27359fd6 | 1270 | dax_unlock_page(page, cookie); |
6100e34b DW |
1271 | out: |
1272 | /* drop pgmap ref acquired in caller */ | |
1273 | put_dev_pagemap(pgmap); | |
1274 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
1275 | return rc; | |
1276 | } | |
1277 | ||
cd42f4a3 TL |
1278 | /** |
1279 | * memory_failure - Handle memory failure of a page. | |
1280 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
1281 | * @flags: fine tune action taken |
1282 | * | |
1283 | * This function is called by the low level machine check code | |
1284 | * of an architecture when it detects hardware memory corruption | |
1285 | * of a page. It tries its best to recover, which includes | |
1286 | * dropping pages, killing processes etc. | |
1287 | * | |
1288 | * The function is primarily of use for corruptions that | |
1289 | * happen outside the current execution context (e.g. when | |
1290 | * detected by a background scrubber) | |
1291 | * | |
1292 | * Must run in process context (e.g. a work queue) with interrupts | |
1293 | * enabled and no spinlocks hold. | |
1294 | */ | |
83b57531 | 1295 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 1296 | { |
6a46079c | 1297 | struct page *p; |
7af446a8 | 1298 | struct page *hpage; |
415c64c1 | 1299 | struct page *orig_head; |
6100e34b | 1300 | struct dev_pagemap *pgmap; |
6a46079c | 1301 | int res; |
524fca1e | 1302 | unsigned long page_flags; |
6a46079c AK |
1303 | |
1304 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 1305 | panic("Memory failure on page %lx", pfn); |
6a46079c | 1306 | |
96c804a6 DH |
1307 | p = pfn_to_online_page(pfn); |
1308 | if (!p) { | |
1309 | if (pfn_valid(pfn)) { | |
1310 | pgmap = get_dev_pagemap(pfn, NULL); | |
1311 | if (pgmap) | |
1312 | return memory_failure_dev_pagemap(pfn, flags, | |
1313 | pgmap); | |
1314 | } | |
495367c0 CY |
1315 | pr_err("Memory failure: %#lx: memory outside kernel control\n", |
1316 | pfn); | |
a7560fc8 | 1317 | return -ENXIO; |
6a46079c AK |
1318 | } |
1319 | ||
761ad8d7 | 1320 | if (PageHuge(p)) |
83b57531 | 1321 | return memory_failure_hugetlb(pfn, flags); |
6a46079c | 1322 | if (TestSetPageHWPoison(p)) { |
495367c0 CY |
1323 | pr_err("Memory failure: %#lx: already hardware poisoned\n", |
1324 | pfn); | |
6a46079c AK |
1325 | return 0; |
1326 | } | |
1327 | ||
761ad8d7 | 1328 | orig_head = hpage = compound_head(p); |
b37ff71c | 1329 | num_poisoned_pages_inc(); |
6a46079c AK |
1330 | |
1331 | /* | |
1332 | * We need/can do nothing about count=0 pages. | |
1333 | * 1) it's a free page, and therefore in safe hand: | |
1334 | * prep_new_page() will be the gate keeper. | |
761ad8d7 | 1335 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
1336 | * Implies some kernel user: cannot stop them from |
1337 | * R/W the page; let's pray that the page has been | |
1338 | * used and will be freed some time later. | |
1339 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 1340 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 1341 | */ |
ead07f6a | 1342 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) { |
8d22ba1b | 1343 | if (is_free_buddy_page(p)) { |
cc637b17 | 1344 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); |
8d22ba1b WF |
1345 | return 0; |
1346 | } else { | |
cc637b17 | 1347 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); |
8d22ba1b WF |
1348 | return -EBUSY; |
1349 | } | |
6a46079c AK |
1350 | } |
1351 | ||
761ad8d7 | 1352 | if (PageTransHuge(hpage)) { |
694bf0b0 | 1353 | if (try_to_split_thp_page(p, "Memory Failure") < 0) |
415c64c1 | 1354 | return -EBUSY; |
415c64c1 | 1355 | VM_BUG_ON_PAGE(!page_count(p), p); |
415c64c1 NH |
1356 | } |
1357 | ||
e43c3afb WF |
1358 | /* |
1359 | * We ignore non-LRU pages for good reasons. | |
1360 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 1361 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
1362 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
1363 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1364 | * walked by the page reclaim code, however that's not a big loss. | |
1365 | */ | |
8bcb74de NH |
1366 | shake_page(p, 0); |
1367 | /* shake_page could have turned it free. */ | |
1368 | if (!PageLRU(p) && is_free_buddy_page(p)) { | |
1369 | if (flags & MF_COUNT_INCREASED) | |
1370 | action_result(pfn, MF_MSG_BUDDY, MF_DELAYED); | |
1371 | else | |
1372 | action_result(pfn, MF_MSG_BUDDY_2ND, MF_DELAYED); | |
1373 | return 0; | |
e43c3afb | 1374 | } |
e43c3afb | 1375 | |
761ad8d7 | 1376 | lock_page(p); |
847ce401 | 1377 | |
f37d4298 AK |
1378 | /* |
1379 | * The page could have changed compound pages during the locking. | |
1380 | * If this happens just bail out. | |
1381 | */ | |
415c64c1 | 1382 | if (PageCompound(p) && compound_head(p) != orig_head) { |
cc637b17 | 1383 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 AK |
1384 | res = -EBUSY; |
1385 | goto out; | |
1386 | } | |
1387 | ||
524fca1e NH |
1388 | /* |
1389 | * We use page flags to determine what action should be taken, but | |
1390 | * the flags can be modified by the error containment action. One | |
1391 | * example is an mlocked page, where PG_mlocked is cleared by | |
1392 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1393 | * correctly, we save a copy of the page flags at this time. | |
1394 | */ | |
7d9d46ac | 1395 | page_flags = p->flags; |
524fca1e | 1396 | |
847ce401 WF |
1397 | /* |
1398 | * unpoison always clear PG_hwpoison inside page lock | |
1399 | */ | |
1400 | if (!PageHWPoison(p)) { | |
495367c0 | 1401 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); |
b37ff71c | 1402 | num_poisoned_pages_dec(); |
761ad8d7 | 1403 | unlock_page(p); |
dd6e2402 | 1404 | put_page(p); |
a09233f3 | 1405 | return 0; |
847ce401 | 1406 | } |
7c116f2b WF |
1407 | if (hwpoison_filter(p)) { |
1408 | if (TestClearPageHWPoison(p)) | |
b37ff71c | 1409 | num_poisoned_pages_dec(); |
761ad8d7 | 1410 | unlock_page(p); |
dd6e2402 | 1411 | put_page(p); |
7c116f2b WF |
1412 | return 0; |
1413 | } | |
847ce401 | 1414 | |
761ad8d7 | 1415 | if (!PageTransTail(p) && !PageLRU(p)) |
0bc1f8b0 CY |
1416 | goto identify_page_state; |
1417 | ||
6edd6cc6 NH |
1418 | /* |
1419 | * It's very difficult to mess with pages currently under IO | |
1420 | * and in many cases impossible, so we just avoid it here. | |
1421 | */ | |
6a46079c AK |
1422 | wait_on_page_writeback(p); |
1423 | ||
1424 | /* | |
1425 | * Now take care of user space mappings. | |
e64a782f | 1426 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
6a46079c | 1427 | */ |
1b473bec | 1428 | if (!hwpoison_user_mappings(p, pfn, flags, &p)) { |
cc637b17 | 1429 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 WF |
1430 | res = -EBUSY; |
1431 | goto out; | |
1432 | } | |
6a46079c AK |
1433 | |
1434 | /* | |
1435 | * Torn down by someone else? | |
1436 | */ | |
dc2a1cbf | 1437 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 1438 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 1439 | res = -EBUSY; |
6a46079c AK |
1440 | goto out; |
1441 | } | |
1442 | ||
0bc1f8b0 | 1443 | identify_page_state: |
0348d2eb | 1444 | res = identify_page_state(pfn, p, page_flags); |
6a46079c | 1445 | out: |
761ad8d7 | 1446 | unlock_page(p); |
6a46079c AK |
1447 | return res; |
1448 | } | |
cd42f4a3 | 1449 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1450 | |
ea8f5fb8 HY |
1451 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1452 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1453 | ||
1454 | struct memory_failure_entry { | |
1455 | unsigned long pfn; | |
ea8f5fb8 HY |
1456 | int flags; |
1457 | }; | |
1458 | ||
1459 | struct memory_failure_cpu { | |
1460 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1461 | MEMORY_FAILURE_FIFO_SIZE); | |
1462 | spinlock_t lock; | |
1463 | struct work_struct work; | |
1464 | }; | |
1465 | ||
1466 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1467 | ||
1468 | /** | |
1469 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1470 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
1471 | * @flags: Flags for memory failure handling |
1472 | * | |
1473 | * This function is called by the low level hardware error handler | |
1474 | * when it detects hardware memory corruption of a page. It schedules | |
1475 | * the recovering of error page, including dropping pages, killing | |
1476 | * processes etc. | |
1477 | * | |
1478 | * The function is primarily of use for corruptions that | |
1479 | * happen outside the current execution context (e.g. when | |
1480 | * detected by a background scrubber) | |
1481 | * | |
1482 | * Can run in IRQ context. | |
1483 | */ | |
83b57531 | 1484 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
1485 | { |
1486 | struct memory_failure_cpu *mf_cpu; | |
1487 | unsigned long proc_flags; | |
1488 | struct memory_failure_entry entry = { | |
1489 | .pfn = pfn, | |
ea8f5fb8 HY |
1490 | .flags = flags, |
1491 | }; | |
1492 | ||
1493 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1494 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 1495 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
1496 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
1497 | else | |
8e33a52f | 1498 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
1499 | pfn); |
1500 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1501 | put_cpu_var(memory_failure_cpu); | |
1502 | } | |
1503 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1504 | ||
1505 | static void memory_failure_work_func(struct work_struct *work) | |
1506 | { | |
1507 | struct memory_failure_cpu *mf_cpu; | |
1508 | struct memory_failure_entry entry = { 0, }; | |
1509 | unsigned long proc_flags; | |
1510 | int gotten; | |
1511 | ||
06202231 | 1512 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
ea8f5fb8 HY |
1513 | for (;;) { |
1514 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1515 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1516 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1517 | if (!gotten) | |
1518 | break; | |
cf870c70 | 1519 | if (entry.flags & MF_SOFT_OFFLINE) |
feec24a6 | 1520 | soft_offline_page(entry.pfn, entry.flags); |
cf870c70 | 1521 | else |
83b57531 | 1522 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
1523 | } |
1524 | } | |
1525 | ||
06202231 JM |
1526 | /* |
1527 | * Process memory_failure work queued on the specified CPU. | |
1528 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. | |
1529 | */ | |
1530 | void memory_failure_queue_kick(int cpu) | |
1531 | { | |
1532 | struct memory_failure_cpu *mf_cpu; | |
1533 | ||
1534 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1535 | cancel_work_sync(&mf_cpu->work); | |
1536 | memory_failure_work_func(&mf_cpu->work); | |
1537 | } | |
1538 | ||
ea8f5fb8 HY |
1539 | static int __init memory_failure_init(void) |
1540 | { | |
1541 | struct memory_failure_cpu *mf_cpu; | |
1542 | int cpu; | |
1543 | ||
1544 | for_each_possible_cpu(cpu) { | |
1545 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1546 | spin_lock_init(&mf_cpu->lock); | |
1547 | INIT_KFIFO(mf_cpu->fifo); | |
1548 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1549 | } | |
1550 | ||
1551 | return 0; | |
1552 | } | |
1553 | core_initcall(memory_failure_init); | |
1554 | ||
a5f65109 NH |
1555 | #define unpoison_pr_info(fmt, pfn, rs) \ |
1556 | ({ \ | |
1557 | if (__ratelimit(rs)) \ | |
1558 | pr_info(fmt, pfn); \ | |
1559 | }) | |
1560 | ||
847ce401 WF |
1561 | /** |
1562 | * unpoison_memory - Unpoison a previously poisoned page | |
1563 | * @pfn: Page number of the to be unpoisoned page | |
1564 | * | |
1565 | * Software-unpoison a page that has been poisoned by | |
1566 | * memory_failure() earlier. | |
1567 | * | |
1568 | * This is only done on the software-level, so it only works | |
1569 | * for linux injected failures, not real hardware failures | |
1570 | * | |
1571 | * Returns 0 for success, otherwise -errno. | |
1572 | */ | |
1573 | int unpoison_memory(unsigned long pfn) | |
1574 | { | |
1575 | struct page *page; | |
1576 | struct page *p; | |
1577 | int freeit = 0; | |
a5f65109 NH |
1578 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
1579 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
1580 | |
1581 | if (!pfn_valid(pfn)) | |
1582 | return -ENXIO; | |
1583 | ||
1584 | p = pfn_to_page(pfn); | |
1585 | page = compound_head(p); | |
1586 | ||
1587 | if (!PageHWPoison(p)) { | |
495367c0 | 1588 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 1589 | pfn, &unpoison_rs); |
847ce401 WF |
1590 | return 0; |
1591 | } | |
1592 | ||
230ac719 | 1593 | if (page_count(page) > 1) { |
495367c0 | 1594 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 1595 | pfn, &unpoison_rs); |
230ac719 NH |
1596 | return 0; |
1597 | } | |
1598 | ||
1599 | if (page_mapped(page)) { | |
495367c0 | 1600 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 1601 | pfn, &unpoison_rs); |
230ac719 NH |
1602 | return 0; |
1603 | } | |
1604 | ||
1605 | if (page_mapping(page)) { | |
495367c0 | 1606 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 1607 | pfn, &unpoison_rs); |
230ac719 NH |
1608 | return 0; |
1609 | } | |
1610 | ||
0cea3fdc WL |
1611 | /* |
1612 | * unpoison_memory() can encounter thp only when the thp is being | |
1613 | * worked by memory_failure() and the page lock is not held yet. | |
1614 | * In such case, we yield to memory_failure() and make unpoison fail. | |
1615 | */ | |
e76d30e2 | 1616 | if (!PageHuge(page) && PageTransHuge(page)) { |
495367c0 | 1617 | unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n", |
a5f65109 | 1618 | pfn, &unpoison_rs); |
ead07f6a | 1619 | return 0; |
0cea3fdc WL |
1620 | } |
1621 | ||
ead07f6a | 1622 | if (!get_hwpoison_page(p)) { |
847ce401 | 1623 | if (TestClearPageHWPoison(p)) |
8e30456b | 1624 | num_poisoned_pages_dec(); |
495367c0 | 1625 | unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n", |
a5f65109 | 1626 | pfn, &unpoison_rs); |
847ce401 WF |
1627 | return 0; |
1628 | } | |
1629 | ||
7eaceacc | 1630 | lock_page(page); |
847ce401 WF |
1631 | /* |
1632 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1633 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1634 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1635 | * the free buddy page pool. | |
1636 | */ | |
c9fbdd5f | 1637 | if (TestClearPageHWPoison(page)) { |
495367c0 | 1638 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", |
a5f65109 | 1639 | pfn, &unpoison_rs); |
b37ff71c | 1640 | num_poisoned_pages_dec(); |
847ce401 WF |
1641 | freeit = 1; |
1642 | } | |
1643 | unlock_page(page); | |
1644 | ||
dd6e2402 | 1645 | put_page(page); |
3ba5eebc | 1646 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) |
dd6e2402 | 1647 | put_page(page); |
847ce401 WF |
1648 | |
1649 | return 0; | |
1650 | } | |
1651 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 1652 | |
666feb21 | 1653 | static struct page *new_page(struct page *p, unsigned long private) |
facb6011 | 1654 | { |
19fc7bed JK |
1655 | struct migration_target_control mtc = { |
1656 | .nid = page_to_nid(p), | |
1657 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
1658 | }; | |
94310cbc | 1659 | |
19fc7bed | 1660 | return alloc_migration_target(p, (unsigned long)&mtc); |
facb6011 AK |
1661 | } |
1662 | ||
1663 | /* | |
1664 | * Safely get reference count of an arbitrary page. | |
1665 | * Returns 0 for a free page, -EIO for a zero refcount page | |
1666 | * that is not free, and 1 for any other page type. | |
1667 | * For 1 the page is returned with increased page count, otherwise not. | |
1668 | */ | |
af8fae7c | 1669 | static int __get_any_page(struct page *p, unsigned long pfn, int flags) |
facb6011 AK |
1670 | { |
1671 | int ret; | |
1672 | ||
1673 | if (flags & MF_COUNT_INCREASED) | |
1674 | return 1; | |
1675 | ||
d950b958 NH |
1676 | /* |
1677 | * When the target page is a free hugepage, just remove it | |
1678 | * from free hugepage list. | |
1679 | */ | |
ead07f6a | 1680 | if (!get_hwpoison_page(p)) { |
d950b958 | 1681 | if (PageHuge(p)) { |
71dd0b8a | 1682 | pr_info("%s: %#lx free huge page\n", __func__, pfn); |
af8fae7c | 1683 | ret = 0; |
d950b958 | 1684 | } else if (is_free_buddy_page(p)) { |
71dd0b8a | 1685 | pr_info("%s: %#lx free buddy page\n", __func__, pfn); |
facb6011 AK |
1686 | ret = 0; |
1687 | } else { | |
71dd0b8a BP |
1688 | pr_info("%s: %#lx: unknown zero refcount page type %lx\n", |
1689 | __func__, pfn, p->flags); | |
facb6011 AK |
1690 | ret = -EIO; |
1691 | } | |
1692 | } else { | |
1693 | /* Not a free page */ | |
1694 | ret = 1; | |
1695 | } | |
facb6011 AK |
1696 | return ret; |
1697 | } | |
1698 | ||
af8fae7c NH |
1699 | static int get_any_page(struct page *page, unsigned long pfn, int flags) |
1700 | { | |
1701 | int ret = __get_any_page(page, pfn, flags); | |
1702 | ||
85fbe5d1 YX |
1703 | if (ret == 1 && !PageHuge(page) && |
1704 | !PageLRU(page) && !__PageMovable(page)) { | |
af8fae7c NH |
1705 | /* |
1706 | * Try to free it. | |
1707 | */ | |
dd6e2402 | 1708 | put_page(page); |
af8fae7c NH |
1709 | shake_page(page, 1); |
1710 | ||
1711 | /* | |
1712 | * Did it turn free? | |
1713 | */ | |
1714 | ret = __get_any_page(page, pfn, 0); | |
d96b339f | 1715 | if (ret == 1 && !PageLRU(page)) { |
4f32be67 | 1716 | /* Drop page reference which is from __get_any_page() */ |
dd6e2402 | 1717 | put_page(page); |
82a2481e AK |
1718 | pr_info("soft_offline: %#lx: unknown non LRU page type %lx (%pGp)\n", |
1719 | pfn, page->flags, &page->flags); | |
af8fae7c NH |
1720 | return -EIO; |
1721 | } | |
1722 | } | |
1723 | return ret; | |
1724 | } | |
1725 | ||
d950b958 NH |
1726 | static int soft_offline_huge_page(struct page *page, int flags) |
1727 | { | |
1728 | int ret; | |
1729 | unsigned long pfn = page_to_pfn(page); | |
1730 | struct page *hpage = compound_head(page); | |
b8ec1cee | 1731 | LIST_HEAD(pagelist); |
d950b958 | 1732 | |
af8fae7c NH |
1733 | /* |
1734 | * This double-check of PageHWPoison is to avoid the race with | |
1735 | * memory_failure(). See also comment in __soft_offline_page(). | |
1736 | */ | |
1737 | lock_page(hpage); | |
0ebff32c | 1738 | if (PageHWPoison(hpage)) { |
af8fae7c | 1739 | unlock_page(hpage); |
dd6e2402 | 1740 | put_page(hpage); |
0ebff32c | 1741 | pr_info("soft offline: %#lx hugepage already poisoned\n", pfn); |
af8fae7c | 1742 | return -EBUSY; |
0ebff32c | 1743 | } |
af8fae7c | 1744 | unlock_page(hpage); |
d950b958 | 1745 | |
bcc54222 | 1746 | ret = isolate_huge_page(hpage, &pagelist); |
03613808 WL |
1747 | /* |
1748 | * get_any_page() and isolate_huge_page() takes a refcount each, | |
1749 | * so need to drop one here. | |
1750 | */ | |
dd6e2402 | 1751 | put_page(hpage); |
03613808 | 1752 | if (!ret) { |
bcc54222 NH |
1753 | pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn); |
1754 | return -EBUSY; | |
1755 | } | |
1756 | ||
68711a74 | 1757 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, |
b8ec1cee | 1758 | MIGRATE_SYNC, MR_MEMORY_FAILURE); |
d950b958 | 1759 | if (ret) { |
b6b18aa8 | 1760 | pr_info("soft offline: %#lx: hugepage migration failed %d, type %lx (%pGp)\n", |
82a2481e | 1761 | pfn, ret, page->flags, &page->flags); |
30809f55 PA |
1762 | if (!list_empty(&pagelist)) |
1763 | putback_movable_pages(&pagelist); | |
b8ec1cee NH |
1764 | if (ret > 0) |
1765 | ret = -EIO; | |
af8fae7c | 1766 | } else { |
6bc9b564 NH |
1767 | /* |
1768 | * We set PG_hwpoison only when the migration source hugepage | |
1769 | * was successfully dissolved, because otherwise hwpoisoned | |
1770 | * hugepage remains on free hugepage list, then userspace will | |
1771 | * find it as SIGBUS by allocation failure. That's not expected | |
1772 | * in soft-offlining. | |
1773 | */ | |
1774 | ret = dissolve_free_huge_page(page); | |
1775 | if (!ret) { | |
1776 | if (set_hwpoison_free_buddy_page(page)) | |
1777 | num_poisoned_pages_inc(); | |
b38e5962 NH |
1778 | else |
1779 | ret = -EBUSY; | |
6bc9b564 | 1780 | } |
d950b958 | 1781 | } |
d950b958 NH |
1782 | return ret; |
1783 | } | |
1784 | ||
af8fae7c NH |
1785 | static int __soft_offline_page(struct page *page, int flags) |
1786 | { | |
1787 | int ret; | |
1788 | unsigned long pfn = page_to_pfn(page); | |
facb6011 | 1789 | |
facb6011 | 1790 | /* |
af8fae7c NH |
1791 | * Check PageHWPoison again inside page lock because PageHWPoison |
1792 | * is set by memory_failure() outside page lock. Note that | |
1793 | * memory_failure() also double-checks PageHWPoison inside page lock, | |
1794 | * so there's no race between soft_offline_page() and memory_failure(). | |
facb6011 | 1795 | */ |
0ebff32c XQ |
1796 | lock_page(page); |
1797 | wait_on_page_writeback(page); | |
af8fae7c NH |
1798 | if (PageHWPoison(page)) { |
1799 | unlock_page(page); | |
dd6e2402 | 1800 | put_page(page); |
af8fae7c NH |
1801 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
1802 | return -EBUSY; | |
1803 | } | |
facb6011 AK |
1804 | /* |
1805 | * Try to invalidate first. This should work for | |
1806 | * non dirty unmapped page cache pages. | |
1807 | */ | |
1808 | ret = invalidate_inode_page(page); | |
1809 | unlock_page(page); | |
facb6011 | 1810 | /* |
facb6011 AK |
1811 | * RED-PEN would be better to keep it isolated here, but we |
1812 | * would need to fix isolation locking first. | |
1813 | */ | |
facb6011 | 1814 | if (ret == 1) { |
dd6e2402 | 1815 | put_page(page); |
fb46e735 | 1816 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
af8fae7c | 1817 | SetPageHWPoison(page); |
8e30456b | 1818 | num_poisoned_pages_inc(); |
af8fae7c | 1819 | return 0; |
facb6011 AK |
1820 | } |
1821 | ||
1822 | /* | |
1823 | * Simple invalidation didn't work. | |
1824 | * Try to migrate to a new page instead. migrate.c | |
1825 | * handles a large number of cases for us. | |
1826 | */ | |
85fbe5d1 YX |
1827 | if (PageLRU(page)) |
1828 | ret = isolate_lru_page(page); | |
1829 | else | |
1830 | ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); | |
bd486285 KK |
1831 | /* |
1832 | * Drop page reference which is came from get_any_page() | |
1833 | * successful isolate_lru_page() already took another one. | |
1834 | */ | |
dd6e2402 | 1835 | put_page(page); |
facb6011 AK |
1836 | if (!ret) { |
1837 | LIST_HEAD(pagelist); | |
85fbe5d1 YX |
1838 | /* |
1839 | * After isolated lru page, the PageLRU will be cleared, | |
1840 | * so use !__PageMovable instead for LRU page's mapping | |
1841 | * cannot have PAGE_MAPPING_MOVABLE. | |
1842 | */ | |
1843 | if (!__PageMovable(page)) | |
1844 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
9de4f22a | 1845 | page_is_file_lru(page)); |
facb6011 | 1846 | list_add(&page->lru, &pagelist); |
68711a74 | 1847 | ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL, |
9c620e2b | 1848 | MIGRATE_SYNC, MR_MEMORY_FAILURE); |
facb6011 | 1849 | if (ret) { |
85fbe5d1 YX |
1850 | if (!list_empty(&pagelist)) |
1851 | putback_movable_pages(&pagelist); | |
59c82b70 | 1852 | |
82a2481e AK |
1853 | pr_info("soft offline: %#lx: migration failed %d, type %lx (%pGp)\n", |
1854 | pfn, ret, page->flags, &page->flags); | |
facb6011 AK |
1855 | if (ret > 0) |
1856 | ret = -EIO; | |
1857 | } | |
1858 | } else { | |
82a2481e AK |
1859 | pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx (%pGp)\n", |
1860 | pfn, ret, page_count(page), page->flags, &page->flags); | |
facb6011 | 1861 | } |
facb6011 AK |
1862 | return ret; |
1863 | } | |
86e05773 | 1864 | |
acc14dc4 NH |
1865 | static int soft_offline_in_use_page(struct page *page, int flags) |
1866 | { | |
1867 | int ret; | |
d4ae9916 | 1868 | int mt; |
acc14dc4 NH |
1869 | struct page *hpage = compound_head(page); |
1870 | ||
694bf0b0 OS |
1871 | if (!PageHuge(page) && PageTransHuge(hpage)) |
1872 | if (try_to_split_thp_page(page, "soft offline") < 0) | |
acc14dc4 | 1873 | return -EBUSY; |
acc14dc4 | 1874 | |
d4ae9916 NH |
1875 | /* |
1876 | * Setting MIGRATE_ISOLATE here ensures that the page will be linked | |
1877 | * to free list immediately (not via pcplist) when released after | |
1878 | * successful page migration. Otherwise we can't guarantee that the | |
1879 | * page is really free after put_page() returns, so | |
1880 | * set_hwpoison_free_buddy_page() highly likely fails. | |
1881 | */ | |
1882 | mt = get_pageblock_migratetype(page); | |
1883 | set_pageblock_migratetype(page, MIGRATE_ISOLATE); | |
acc14dc4 NH |
1884 | if (PageHuge(page)) |
1885 | ret = soft_offline_huge_page(page, flags); | |
1886 | else | |
1887 | ret = __soft_offline_page(page, flags); | |
d4ae9916 | 1888 | set_pageblock_migratetype(page, mt); |
acc14dc4 NH |
1889 | return ret; |
1890 | } | |
1891 | ||
d4ae9916 | 1892 | static int soft_offline_free_page(struct page *page) |
acc14dc4 | 1893 | { |
06be6ff3 | 1894 | int rc = -EBUSY; |
acc14dc4 | 1895 | |
06be6ff3 OS |
1896 | if (!dissolve_free_huge_page(page) && take_page_off_buddy(page)) { |
1897 | page_handle_poison(page); | |
1898 | rc = 0; | |
d4ae9916 | 1899 | } |
06be6ff3 | 1900 | |
d4ae9916 | 1901 | return rc; |
acc14dc4 NH |
1902 | } |
1903 | ||
86e05773 WL |
1904 | /** |
1905 | * soft_offline_page - Soft offline a page. | |
feec24a6 | 1906 | * @pfn: pfn to soft-offline |
86e05773 WL |
1907 | * @flags: flags. Same as memory_failure(). |
1908 | * | |
1909 | * Returns 0 on success, otherwise negated errno. | |
1910 | * | |
1911 | * Soft offline a page, by migration or invalidation, | |
1912 | * without killing anything. This is for the case when | |
1913 | * a page is not corrupted yet (so it's still valid to access), | |
1914 | * but has had a number of corrected errors and is better taken | |
1915 | * out. | |
1916 | * | |
1917 | * The actual policy on when to do that is maintained by | |
1918 | * user space. | |
1919 | * | |
1920 | * This should never impact any application or cause data loss, | |
1921 | * however it might take some time. | |
1922 | * | |
1923 | * This is not a 100% solution for all memory, but tries to be | |
1924 | * ``good enough'' for the majority of memory. | |
1925 | */ | |
feec24a6 | 1926 | int soft_offline_page(unsigned long pfn, int flags) |
86e05773 WL |
1927 | { |
1928 | int ret; | |
feec24a6 | 1929 | struct page *page; |
86e05773 | 1930 | |
feec24a6 NH |
1931 | if (!pfn_valid(pfn)) |
1932 | return -ENXIO; | |
1933 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ | |
1934 | page = pfn_to_online_page(pfn); | |
1935 | if (!page) | |
86a66810 | 1936 | return -EIO; |
86a66810 | 1937 | |
86e05773 WL |
1938 | if (PageHWPoison(page)) { |
1939 | pr_info("soft offline: %#lx page already poisoned\n", pfn); | |
1e0e635b | 1940 | if (flags & MF_COUNT_INCREASED) |
dd6e2402 | 1941 | put_page(page); |
86e05773 WL |
1942 | return -EBUSY; |
1943 | } | |
86e05773 | 1944 | |
bfc8c901 | 1945 | get_online_mems(); |
86e05773 | 1946 | ret = get_any_page(page, pfn, flags); |
bfc8c901 | 1947 | put_online_mems(); |
4e41a30c | 1948 | |
acc14dc4 NH |
1949 | if (ret > 0) |
1950 | ret = soft_offline_in_use_page(page, flags); | |
1951 | else if (ret == 0) | |
d4ae9916 | 1952 | ret = soft_offline_free_page(page); |
4e41a30c | 1953 | |
86e05773 WL |
1954 | return ret; |
1955 | } |