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> |
a3f5d80e | 59 | #include <linux/pagewalk.h> |
6a46079c | 60 | #include "internal.h" |
97f0b134 | 61 | #include "ras/ras_event.h" |
6a46079c AK |
62 | |
63 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
64 | ||
65 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
66 | ||
293c07e3 | 67 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 68 | |
510d25c9 NH |
69 | static bool __page_handle_poison(struct page *page) |
70 | { | |
71 | bool ret; | |
72 | ||
73 | zone_pcp_disable(page_zone(page)); | |
74 | ret = dissolve_free_huge_page(page); | |
75 | if (!ret) | |
76 | ret = take_page_off_buddy(page); | |
77 | zone_pcp_enable(page_zone(page)); | |
78 | ||
79 | return ret; | |
80 | } | |
81 | ||
6b9a217e | 82 | static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release) |
06be6ff3 | 83 | { |
6b9a217e OS |
84 | if (hugepage_or_freepage) { |
85 | /* | |
86 | * Doing this check for free pages is also fine since dissolve_free_huge_page | |
87 | * returns 0 for non-hugetlb pages as well. | |
88 | */ | |
510d25c9 | 89 | if (!__page_handle_poison(page)) |
6b9a217e OS |
90 | /* |
91 | * We could fail to take off the target page from buddy | |
f0953a1b | 92 | * for example due to racy page allocation, but that's |
6b9a217e OS |
93 | * acceptable because soft-offlined page is not broken |
94 | * and if someone really want to use it, they should | |
95 | * take it. | |
96 | */ | |
97 | return false; | |
98 | } | |
99 | ||
06be6ff3 | 100 | SetPageHWPoison(page); |
79f5f8fa OS |
101 | if (release) |
102 | put_page(page); | |
06be6ff3 OS |
103 | page_ref_inc(page); |
104 | num_poisoned_pages_inc(); | |
6b9a217e OS |
105 | |
106 | return true; | |
06be6ff3 OS |
107 | } |
108 | ||
27df5068 AK |
109 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
110 | ||
1bfe5feb | 111 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
112 | u32 hwpoison_filter_dev_major = ~0U; |
113 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
114 | u64 hwpoison_filter_flags_mask; |
115 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 116 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
117 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
118 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
119 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
120 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
121 | |
122 | static int hwpoison_filter_dev(struct page *p) | |
123 | { | |
124 | struct address_space *mapping; | |
125 | dev_t dev; | |
126 | ||
127 | if (hwpoison_filter_dev_major == ~0U && | |
128 | hwpoison_filter_dev_minor == ~0U) | |
129 | return 0; | |
130 | ||
131 | /* | |
1c80b990 | 132 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
133 | */ |
134 | if (PageSlab(p)) | |
135 | return -EINVAL; | |
136 | ||
137 | mapping = page_mapping(p); | |
138 | if (mapping == NULL || mapping->host == NULL) | |
139 | return -EINVAL; | |
140 | ||
141 | dev = mapping->host->i_sb->s_dev; | |
142 | if (hwpoison_filter_dev_major != ~0U && | |
143 | hwpoison_filter_dev_major != MAJOR(dev)) | |
144 | return -EINVAL; | |
145 | if (hwpoison_filter_dev_minor != ~0U && | |
146 | hwpoison_filter_dev_minor != MINOR(dev)) | |
147 | return -EINVAL; | |
148 | ||
149 | return 0; | |
150 | } | |
151 | ||
478c5ffc WF |
152 | static int hwpoison_filter_flags(struct page *p) |
153 | { | |
154 | if (!hwpoison_filter_flags_mask) | |
155 | return 0; | |
156 | ||
157 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
158 | hwpoison_filter_flags_value) | |
159 | return 0; | |
160 | else | |
161 | return -EINVAL; | |
162 | } | |
163 | ||
4fd466eb AK |
164 | /* |
165 | * This allows stress tests to limit test scope to a collection of tasks | |
166 | * by putting them under some memcg. This prevents killing unrelated/important | |
167 | * processes such as /sbin/init. Note that the target task may share clean | |
168 | * pages with init (eg. libc text), which is harmless. If the target task | |
169 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
170 | * is also included in the memcg. At last, due to race conditions this filter | |
171 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
172 | * a freed page. | |
173 | */ | |
94a59fb3 | 174 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
175 | u64 hwpoison_filter_memcg; |
176 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
177 | static int hwpoison_filter_task(struct page *p) | |
178 | { | |
4fd466eb AK |
179 | if (!hwpoison_filter_memcg) |
180 | return 0; | |
181 | ||
94a59fb3 | 182 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
183 | return -EINVAL; |
184 | ||
185 | return 0; | |
186 | } | |
187 | #else | |
188 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
189 | #endif | |
190 | ||
7c116f2b WF |
191 | int hwpoison_filter(struct page *p) |
192 | { | |
1bfe5feb HL |
193 | if (!hwpoison_filter_enable) |
194 | return 0; | |
195 | ||
7c116f2b WF |
196 | if (hwpoison_filter_dev(p)) |
197 | return -EINVAL; | |
198 | ||
478c5ffc WF |
199 | if (hwpoison_filter_flags(p)) |
200 | return -EINVAL; | |
201 | ||
4fd466eb AK |
202 | if (hwpoison_filter_task(p)) |
203 | return -EINVAL; | |
204 | ||
7c116f2b WF |
205 | return 0; |
206 | } | |
27df5068 AK |
207 | #else |
208 | int hwpoison_filter(struct page *p) | |
209 | { | |
210 | return 0; | |
211 | } | |
212 | #endif | |
213 | ||
7c116f2b WF |
214 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
215 | ||
ae1139ec DW |
216 | /* |
217 | * Kill all processes that have a poisoned page mapped and then isolate | |
218 | * the page. | |
219 | * | |
220 | * General strategy: | |
221 | * Find all processes having the page mapped and kill them. | |
222 | * But we keep a page reference around so that the page is not | |
223 | * actually freed yet. | |
224 | * Then stash the page away | |
225 | * | |
226 | * There's no convenient way to get back to mapped processes | |
227 | * from the VMAs. So do a brute-force search over all | |
228 | * running processes. | |
229 | * | |
230 | * Remember that machine checks are not common (or rather | |
231 | * if they are common you have other problems), so this shouldn't | |
232 | * be a performance issue. | |
233 | * | |
234 | * Also there are some races possible while we get from the | |
235 | * error detection to actually handle it. | |
236 | */ | |
237 | ||
238 | struct to_kill { | |
239 | struct list_head nd; | |
240 | struct task_struct *tsk; | |
241 | unsigned long addr; | |
242 | short size_shift; | |
ae1139ec DW |
243 | }; |
244 | ||
6a46079c | 245 | /* |
7329bbeb TL |
246 | * Send all the processes who have the page mapped a signal. |
247 | * ``action optional'' if they are not immediately affected by the error | |
248 | * ``action required'' if error happened in current execution context | |
6a46079c | 249 | */ |
ae1139ec | 250 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 251 | { |
ae1139ec DW |
252 | struct task_struct *t = tk->tsk; |
253 | short addr_lsb = tk->size_shift; | |
872e9a20 | 254 | int ret = 0; |
6a46079c | 255 | |
03151c6e | 256 | pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", |
872e9a20 | 257 | pfn, t->comm, t->pid); |
7329bbeb | 258 | |
872e9a20 | 259 | if (flags & MF_ACTION_REQUIRED) { |
30c9cf49 AY |
260 | if (t == current) |
261 | ret = force_sig_mceerr(BUS_MCEERR_AR, | |
872e9a20 | 262 | (void __user *)tk->addr, addr_lsb); |
30c9cf49 AY |
263 | else |
264 | /* Signal other processes sharing the page if they have PF_MCE_EARLY set. */ | |
265 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, | |
266 | addr_lsb, t); | |
7329bbeb TL |
267 | } else { |
268 | /* | |
269 | * Don't use force here, it's convenient if the signal | |
270 | * can be temporarily blocked. | |
271 | * This could cause a loop when the user sets SIGBUS | |
272 | * to SIG_IGN, but hopefully no one will do that? | |
273 | */ | |
ae1139ec | 274 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
c0f45555 | 275 | addr_lsb, t); /* synchronous? */ |
7329bbeb | 276 | } |
6a46079c | 277 | if (ret < 0) |
495367c0 | 278 | pr_info("Memory failure: Error sending signal to %s:%d: %d\n", |
1170532b | 279 | t->comm, t->pid, ret); |
6a46079c AK |
280 | return ret; |
281 | } | |
282 | ||
588f9ce6 | 283 | /* |
47e431f4 OS |
284 | * Unknown page type encountered. Try to check whether it can turn PageLRU by |
285 | * lru_add_drain_all, or a free page by reclaiming slabs when possible. | |
588f9ce6 | 286 | */ |
facb6011 | 287 | void shake_page(struct page *p, int access) |
588f9ce6 | 288 | { |
8bcb74de NH |
289 | if (PageHuge(p)) |
290 | return; | |
291 | ||
588f9ce6 AK |
292 | if (!PageSlab(p)) { |
293 | lru_add_drain_all(); | |
588f9ce6 AK |
294 | if (PageLRU(p) || is_free_buddy_page(p)) |
295 | return; | |
296 | } | |
facb6011 | 297 | |
588f9ce6 | 298 | /* |
6b4f7799 JW |
299 | * Only call shrink_node_slabs here (which would also shrink |
300 | * other caches) if access is not potentially fatal. | |
588f9ce6 | 301 | */ |
cb731d6c VD |
302 | if (access) |
303 | drop_slab_node(page_to_nid(p)); | |
588f9ce6 AK |
304 | } |
305 | EXPORT_SYMBOL_GPL(shake_page); | |
306 | ||
6100e34b DW |
307 | static unsigned long dev_pagemap_mapping_shift(struct page *page, |
308 | struct vm_area_struct *vma) | |
309 | { | |
310 | unsigned long address = vma_address(page, vma); | |
311 | pgd_t *pgd; | |
312 | p4d_t *p4d; | |
313 | pud_t *pud; | |
314 | pmd_t *pmd; | |
315 | pte_t *pte; | |
316 | ||
317 | pgd = pgd_offset(vma->vm_mm, address); | |
318 | if (!pgd_present(*pgd)) | |
319 | return 0; | |
320 | p4d = p4d_offset(pgd, address); | |
321 | if (!p4d_present(*p4d)) | |
322 | return 0; | |
323 | pud = pud_offset(p4d, address); | |
324 | if (!pud_present(*pud)) | |
325 | return 0; | |
326 | if (pud_devmap(*pud)) | |
327 | return PUD_SHIFT; | |
328 | pmd = pmd_offset(pud, address); | |
329 | if (!pmd_present(*pmd)) | |
330 | return 0; | |
331 | if (pmd_devmap(*pmd)) | |
332 | return PMD_SHIFT; | |
333 | pte = pte_offset_map(pmd, address); | |
334 | if (!pte_present(*pte)) | |
335 | return 0; | |
336 | if (pte_devmap(*pte)) | |
337 | return PAGE_SHIFT; | |
338 | return 0; | |
339 | } | |
6a46079c AK |
340 | |
341 | /* | |
342 | * Failure handling: if we can't find or can't kill a process there's | |
343 | * not much we can do. We just print a message and ignore otherwise. | |
344 | */ | |
345 | ||
346 | /* | |
347 | * Schedule a process for later kill. | |
348 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
6a46079c AK |
349 | */ |
350 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
351 | struct vm_area_struct *vma, | |
996ff7a0 | 352 | struct list_head *to_kill) |
6a46079c AK |
353 | { |
354 | struct to_kill *tk; | |
355 | ||
996ff7a0 JC |
356 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
357 | if (!tk) { | |
358 | pr_err("Memory failure: Out of memory while machine check handling\n"); | |
359 | return; | |
6a46079c | 360 | } |
996ff7a0 | 361 | |
6a46079c | 362 | tk->addr = page_address_in_vma(p, vma); |
6100e34b DW |
363 | if (is_zone_device_page(p)) |
364 | tk->size_shift = dev_pagemap_mapping_shift(p, vma); | |
365 | else | |
75068518 | 366 | tk->size_shift = page_shift(compound_head(p)); |
6a46079c AK |
367 | |
368 | /* | |
3d7fed4a JC |
369 | * Send SIGKILL if "tk->addr == -EFAULT". Also, as |
370 | * "tk->size_shift" is always non-zero for !is_zone_device_page(), | |
371 | * so "tk->size_shift == 0" effectively checks no mapping on | |
372 | * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times | |
373 | * to a process' address space, it's possible not all N VMAs | |
374 | * contain mappings for the page, but at least one VMA does. | |
375 | * Only deliver SIGBUS with payload derived from the VMA that | |
376 | * has a mapping for the page. | |
6a46079c | 377 | */ |
3d7fed4a | 378 | if (tk->addr == -EFAULT) { |
495367c0 | 379 | pr_info("Memory failure: Unable to find user space address %lx in %s\n", |
6a46079c | 380 | page_to_pfn(p), tsk->comm); |
3d7fed4a JC |
381 | } else if (tk->size_shift == 0) { |
382 | kfree(tk); | |
383 | return; | |
6a46079c | 384 | } |
996ff7a0 | 385 | |
6a46079c AK |
386 | get_task_struct(tsk); |
387 | tk->tsk = tsk; | |
388 | list_add_tail(&tk->nd, to_kill); | |
389 | } | |
390 | ||
391 | /* | |
392 | * Kill the processes that have been collected earlier. | |
393 | * | |
394 | * Only do anything when DOIT is set, otherwise just free the list | |
395 | * (this is used for clean pages which do not need killing) | |
396 | * Also when FAIL is set do a force kill because something went | |
397 | * wrong earlier. | |
398 | */ | |
ae1139ec DW |
399 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
400 | unsigned long pfn, int flags) | |
6a46079c AK |
401 | { |
402 | struct to_kill *tk, *next; | |
403 | ||
404 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 405 | if (forcekill) { |
6a46079c | 406 | /* |
af901ca1 | 407 | * In case something went wrong with munmapping |
6a46079c AK |
408 | * make sure the process doesn't catch the |
409 | * signal and then access the memory. Just kill it. | |
6a46079c | 410 | */ |
3d7fed4a | 411 | if (fail || tk->addr == -EFAULT) { |
495367c0 | 412 | pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 413 | pfn, tk->tsk->comm, tk->tsk->pid); |
6376360e NH |
414 | do_send_sig_info(SIGKILL, SEND_SIG_PRIV, |
415 | tk->tsk, PIDTYPE_PID); | |
6a46079c AK |
416 | } |
417 | ||
418 | /* | |
419 | * In theory the process could have mapped | |
420 | * something else on the address in-between. We could | |
421 | * check for that, but we need to tell the | |
422 | * process anyways. | |
423 | */ | |
ae1139ec | 424 | else if (kill_proc(tk, pfn, flags) < 0) |
495367c0 | 425 | pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 426 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c AK |
427 | } |
428 | put_task_struct(tk->tsk); | |
429 | kfree(tk); | |
430 | } | |
431 | } | |
432 | ||
3ba08129 NH |
433 | /* |
434 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
435 | * on behalf of the thread group. Return task_struct of the (first found) | |
436 | * dedicated thread if found, and return NULL otherwise. | |
437 | * | |
438 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
439 | * have to call rcu_read_lock/unlock() in this function. | |
440 | */ | |
441 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 442 | { |
3ba08129 NH |
443 | struct task_struct *t; |
444 | ||
4e018b45 NH |
445 | for_each_thread(tsk, t) { |
446 | if (t->flags & PF_MCE_PROCESS) { | |
447 | if (t->flags & PF_MCE_EARLY) | |
448 | return t; | |
449 | } else { | |
450 | if (sysctl_memory_failure_early_kill) | |
451 | return t; | |
452 | } | |
453 | } | |
3ba08129 NH |
454 | return NULL; |
455 | } | |
456 | ||
457 | /* | |
458 | * Determine whether a given process is "early kill" process which expects | |
459 | * to be signaled when some page under the process is hwpoisoned. | |
460 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
30c9cf49 | 461 | * specified) if the process is "early kill" and otherwise returns NULL. |
03151c6e | 462 | * |
30c9cf49 AY |
463 | * Note that the above is true for Action Optional case. For Action Required |
464 | * case, it's only meaningful to the current thread which need to be signaled | |
465 | * with SIGBUS, this error is Action Optional for other non current | |
466 | * processes sharing the same error page,if the process is "early kill", the | |
467 | * task_struct of the dedicated thread will also be returned. | |
3ba08129 NH |
468 | */ |
469 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
470 | int force_early) | |
471 | { | |
6a46079c | 472 | if (!tsk->mm) |
3ba08129 | 473 | return NULL; |
30c9cf49 AY |
474 | /* |
475 | * Comparing ->mm here because current task might represent | |
476 | * a subthread, while tsk always points to the main thread. | |
477 | */ | |
478 | if (force_early && tsk->mm == current->mm) | |
479 | return current; | |
480 | ||
4e018b45 | 481 | return find_early_kill_thread(tsk); |
6a46079c AK |
482 | } |
483 | ||
484 | /* | |
485 | * Collect processes when the error hit an anonymous page. | |
486 | */ | |
487 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 488 | int force_early) |
6a46079c AK |
489 | { |
490 | struct vm_area_struct *vma; | |
491 | struct task_struct *tsk; | |
492 | struct anon_vma *av; | |
bf181b9f | 493 | pgoff_t pgoff; |
6a46079c | 494 | |
4fc3f1d6 | 495 | av = page_lock_anon_vma_read(page); |
6a46079c | 496 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
497 | return; |
498 | ||
a0f7a756 | 499 | pgoff = page_to_pgoff(page); |
9b679320 | 500 | read_lock(&tasklist_lock); |
6a46079c | 501 | for_each_process (tsk) { |
5beb4930 | 502 | struct anon_vma_chain *vmac; |
3ba08129 | 503 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 504 | |
3ba08129 | 505 | if (!t) |
6a46079c | 506 | continue; |
bf181b9f ML |
507 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
508 | pgoff, pgoff) { | |
5beb4930 | 509 | vma = vmac->vma; |
6a46079c AK |
510 | if (!page_mapped_in_vma(page, vma)) |
511 | continue; | |
3ba08129 | 512 | if (vma->vm_mm == t->mm) |
996ff7a0 | 513 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
514 | } |
515 | } | |
6a46079c | 516 | read_unlock(&tasklist_lock); |
4fc3f1d6 | 517 | page_unlock_anon_vma_read(av); |
6a46079c AK |
518 | } |
519 | ||
520 | /* | |
521 | * Collect processes when the error hit a file mapped page. | |
522 | */ | |
523 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 524 | int force_early) |
6a46079c AK |
525 | { |
526 | struct vm_area_struct *vma; | |
527 | struct task_struct *tsk; | |
6a46079c | 528 | struct address_space *mapping = page->mapping; |
c43bc03d | 529 | pgoff_t pgoff; |
6a46079c | 530 | |
d28eb9c8 | 531 | i_mmap_lock_read(mapping); |
9b679320 | 532 | read_lock(&tasklist_lock); |
c43bc03d | 533 | pgoff = page_to_pgoff(page); |
6a46079c | 534 | for_each_process(tsk) { |
3ba08129 | 535 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 536 | |
3ba08129 | 537 | if (!t) |
6a46079c | 538 | continue; |
6b2dbba8 | 539 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
540 | pgoff) { |
541 | /* | |
542 | * Send early kill signal to tasks where a vma covers | |
543 | * the page but the corrupted page is not necessarily | |
544 | * mapped it in its pte. | |
545 | * Assume applications who requested early kill want | |
546 | * to be informed of all such data corruptions. | |
547 | */ | |
3ba08129 | 548 | if (vma->vm_mm == t->mm) |
996ff7a0 | 549 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
550 | } |
551 | } | |
6a46079c | 552 | read_unlock(&tasklist_lock); |
d28eb9c8 | 553 | i_mmap_unlock_read(mapping); |
6a46079c AK |
554 | } |
555 | ||
556 | /* | |
557 | * Collect the processes who have the corrupted page mapped to kill. | |
6a46079c | 558 | */ |
74614de1 TL |
559 | static void collect_procs(struct page *page, struct list_head *tokill, |
560 | int force_early) | |
6a46079c | 561 | { |
6a46079c AK |
562 | if (!page->mapping) |
563 | return; | |
564 | ||
6a46079c | 565 | if (PageAnon(page)) |
996ff7a0 | 566 | collect_procs_anon(page, tokill, force_early); |
6a46079c | 567 | else |
996ff7a0 | 568 | collect_procs_file(page, tokill, force_early); |
6a46079c AK |
569 | } |
570 | ||
a3f5d80e NH |
571 | struct hwp_walk { |
572 | struct to_kill tk; | |
573 | unsigned long pfn; | |
574 | int flags; | |
575 | }; | |
576 | ||
577 | static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift) | |
578 | { | |
579 | tk->addr = addr; | |
580 | tk->size_shift = shift; | |
581 | } | |
582 | ||
583 | static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift, | |
584 | unsigned long poisoned_pfn, struct to_kill *tk) | |
585 | { | |
586 | unsigned long pfn = 0; | |
587 | ||
588 | if (pte_present(pte)) { | |
589 | pfn = pte_pfn(pte); | |
590 | } else { | |
591 | swp_entry_t swp = pte_to_swp_entry(pte); | |
592 | ||
593 | if (is_hwpoison_entry(swp)) | |
594 | pfn = hwpoison_entry_to_pfn(swp); | |
595 | } | |
596 | ||
597 | if (!pfn || pfn != poisoned_pfn) | |
598 | return 0; | |
599 | ||
600 | set_to_kill(tk, addr, shift); | |
601 | return 1; | |
602 | } | |
603 | ||
604 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
605 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
606 | struct hwp_walk *hwp) | |
607 | { | |
608 | pmd_t pmd = *pmdp; | |
609 | unsigned long pfn; | |
610 | unsigned long hwpoison_vaddr; | |
611 | ||
612 | if (!pmd_present(pmd)) | |
613 | return 0; | |
614 | pfn = pmd_pfn(pmd); | |
615 | if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) { | |
616 | hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT); | |
617 | set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT); | |
618 | return 1; | |
619 | } | |
620 | return 0; | |
621 | } | |
622 | #else | |
623 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
624 | struct hwp_walk *hwp) | |
625 | { | |
626 | return 0; | |
627 | } | |
628 | #endif | |
629 | ||
630 | static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr, | |
631 | unsigned long end, struct mm_walk *walk) | |
632 | { | |
633 | struct hwp_walk *hwp = (struct hwp_walk *)walk->private; | |
634 | int ret = 0; | |
635 | pte_t *ptep; | |
636 | spinlock_t *ptl; | |
637 | ||
638 | ptl = pmd_trans_huge_lock(pmdp, walk->vma); | |
639 | if (ptl) { | |
640 | ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp); | |
641 | spin_unlock(ptl); | |
642 | goto out; | |
643 | } | |
644 | ||
645 | if (pmd_trans_unstable(pmdp)) | |
646 | goto out; | |
647 | ||
648 | ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, addr, &ptl); | |
649 | for (; addr != end; ptep++, addr += PAGE_SIZE) { | |
650 | ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT, | |
651 | hwp->pfn, &hwp->tk); | |
652 | if (ret == 1) | |
653 | break; | |
654 | } | |
655 | pte_unmap_unlock(ptep - 1, ptl); | |
656 | out: | |
657 | cond_resched(); | |
658 | return ret; | |
659 | } | |
660 | ||
661 | #ifdef CONFIG_HUGETLB_PAGE | |
662 | static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask, | |
663 | unsigned long addr, unsigned long end, | |
664 | struct mm_walk *walk) | |
665 | { | |
666 | struct hwp_walk *hwp = (struct hwp_walk *)walk->private; | |
667 | pte_t pte = huge_ptep_get(ptep); | |
668 | struct hstate *h = hstate_vma(walk->vma); | |
669 | ||
670 | return check_hwpoisoned_entry(pte, addr, huge_page_shift(h), | |
671 | hwp->pfn, &hwp->tk); | |
672 | } | |
673 | #else | |
674 | #define hwpoison_hugetlb_range NULL | |
675 | #endif | |
676 | ||
677 | static struct mm_walk_ops hwp_walk_ops = { | |
678 | .pmd_entry = hwpoison_pte_range, | |
679 | .hugetlb_entry = hwpoison_hugetlb_range, | |
680 | }; | |
681 | ||
682 | /* | |
683 | * Sends SIGBUS to the current process with error info. | |
684 | * | |
685 | * This function is intended to handle "Action Required" MCEs on already | |
686 | * hardware poisoned pages. They could happen, for example, when | |
687 | * memory_failure() failed to unmap the error page at the first call, or | |
688 | * when multiple local machine checks happened on different CPUs. | |
689 | * | |
690 | * MCE handler currently has no easy access to the error virtual address, | |
691 | * so this function walks page table to find it. The returned virtual address | |
692 | * is proper in most cases, but it could be wrong when the application | |
693 | * process has multiple entries mapping the error page. | |
694 | */ | |
695 | static int kill_accessing_process(struct task_struct *p, unsigned long pfn, | |
696 | int flags) | |
697 | { | |
698 | int ret; | |
699 | struct hwp_walk priv = { | |
700 | .pfn = pfn, | |
701 | }; | |
702 | priv.tk.tsk = p; | |
703 | ||
704 | mmap_read_lock(p->mm); | |
705 | ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops, | |
706 | (void *)&priv); | |
707 | if (ret == 1 && priv.tk.addr) | |
708 | kill_proc(&priv.tk, pfn, flags); | |
709 | mmap_read_unlock(p->mm); | |
710 | return ret ? -EFAULT : -EHWPOISON; | |
711 | } | |
712 | ||
6a46079c | 713 | static const char *action_name[] = { |
cc637b17 XX |
714 | [MF_IGNORED] = "Ignored", |
715 | [MF_FAILED] = "Failed", | |
716 | [MF_DELAYED] = "Delayed", | |
717 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
718 | }; |
719 | ||
720 | static const char * const action_page_types[] = { | |
cc637b17 XX |
721 | [MF_MSG_KERNEL] = "reserved kernel page", |
722 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
723 | [MF_MSG_SLAB] = "kernel slab page", | |
724 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
725 | [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", | |
726 | [MF_MSG_HUGE] = "huge page", | |
727 | [MF_MSG_FREE_HUGE] = "free huge page", | |
31286a84 | 728 | [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page", |
cc637b17 XX |
729 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", |
730 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
731 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
732 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
733 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
734 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
735 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
736 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
737 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
738 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
739 | [MF_MSG_BUDDY] = "free buddy page", | |
740 | [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", | |
6100e34b | 741 | [MF_MSG_DAX] = "dax page", |
5d1fd5dc | 742 | [MF_MSG_UNSPLIT_THP] = "unsplit thp", |
cc637b17 | 743 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
744 | }; |
745 | ||
dc2a1cbf WF |
746 | /* |
747 | * XXX: It is possible that a page is isolated from LRU cache, | |
748 | * and then kept in swap cache or failed to remove from page cache. | |
749 | * The page count will stop it from being freed by unpoison. | |
750 | * Stress tests should be aware of this memory leak problem. | |
751 | */ | |
752 | static int delete_from_lru_cache(struct page *p) | |
753 | { | |
754 | if (!isolate_lru_page(p)) { | |
755 | /* | |
756 | * Clear sensible page flags, so that the buddy system won't | |
757 | * complain when the page is unpoison-and-freed. | |
758 | */ | |
759 | ClearPageActive(p); | |
760 | ClearPageUnevictable(p); | |
18365225 MH |
761 | |
762 | /* | |
763 | * Poisoned page might never drop its ref count to 0 so we have | |
764 | * to uncharge it manually from its memcg. | |
765 | */ | |
766 | mem_cgroup_uncharge(p); | |
767 | ||
dc2a1cbf WF |
768 | /* |
769 | * drop the page count elevated by isolate_lru_page() | |
770 | */ | |
09cbfeaf | 771 | put_page(p); |
dc2a1cbf WF |
772 | return 0; |
773 | } | |
774 | return -EIO; | |
775 | } | |
776 | ||
78bb9203 NH |
777 | static int truncate_error_page(struct page *p, unsigned long pfn, |
778 | struct address_space *mapping) | |
779 | { | |
780 | int ret = MF_FAILED; | |
781 | ||
782 | if (mapping->a_ops->error_remove_page) { | |
783 | int err = mapping->a_ops->error_remove_page(mapping, p); | |
784 | ||
785 | if (err != 0) { | |
786 | pr_info("Memory failure: %#lx: Failed to punch page: %d\n", | |
787 | pfn, err); | |
788 | } else if (page_has_private(p) && | |
789 | !try_to_release_page(p, GFP_NOIO)) { | |
790 | pr_info("Memory failure: %#lx: failed to release buffers\n", | |
791 | pfn); | |
792 | } else { | |
793 | ret = MF_RECOVERED; | |
794 | } | |
795 | } else { | |
796 | /* | |
797 | * If the file system doesn't support it just invalidate | |
798 | * This fails on dirty or anything with private pages | |
799 | */ | |
800 | if (invalidate_inode_page(p)) | |
801 | ret = MF_RECOVERED; | |
802 | else | |
803 | pr_info("Memory failure: %#lx: Failed to invalidate\n", | |
804 | pfn); | |
805 | } | |
806 | ||
807 | return ret; | |
808 | } | |
809 | ||
6a46079c AK |
810 | /* |
811 | * Error hit kernel page. | |
812 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
813 | * could be more sophisticated. | |
814 | */ | |
815 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c | 816 | { |
ea6d0630 | 817 | unlock_page(p); |
cc637b17 | 818 | return MF_IGNORED; |
6a46079c AK |
819 | } |
820 | ||
821 | /* | |
822 | * Page in unknown state. Do nothing. | |
823 | */ | |
824 | static int me_unknown(struct page *p, unsigned long pfn) | |
825 | { | |
495367c0 | 826 | pr_err("Memory failure: %#lx: Unknown page state\n", pfn); |
ea6d0630 | 827 | unlock_page(p); |
cc637b17 | 828 | return MF_FAILED; |
6a46079c AK |
829 | } |
830 | ||
6a46079c AK |
831 | /* |
832 | * Clean (or cleaned) page cache page. | |
833 | */ | |
834 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
835 | { | |
ea6d0630 | 836 | int ret; |
6a46079c AK |
837 | struct address_space *mapping; |
838 | ||
dc2a1cbf WF |
839 | delete_from_lru_cache(p); |
840 | ||
6a46079c AK |
841 | /* |
842 | * For anonymous pages we're done the only reference left | |
843 | * should be the one m_f() holds. | |
844 | */ | |
ea6d0630 NH |
845 | if (PageAnon(p)) { |
846 | ret = MF_RECOVERED; | |
847 | goto out; | |
848 | } | |
6a46079c AK |
849 | |
850 | /* | |
851 | * Now truncate the page in the page cache. This is really | |
852 | * more like a "temporary hole punch" | |
853 | * Don't do this for block devices when someone else | |
854 | * has a reference, because it could be file system metadata | |
855 | * and that's not safe to truncate. | |
856 | */ | |
857 | mapping = page_mapping(p); | |
858 | if (!mapping) { | |
859 | /* | |
860 | * Page has been teared down in the meanwhile | |
861 | */ | |
ea6d0630 NH |
862 | ret = MF_FAILED; |
863 | goto out; | |
6a46079c AK |
864 | } |
865 | ||
866 | /* | |
867 | * Truncation is a bit tricky. Enable it per file system for now. | |
868 | * | |
869 | * Open: to take i_mutex or not for this? Right now we don't. | |
870 | */ | |
ea6d0630 NH |
871 | ret = truncate_error_page(p, pfn, mapping); |
872 | out: | |
873 | unlock_page(p); | |
874 | return ret; | |
6a46079c AK |
875 | } |
876 | ||
877 | /* | |
549543df | 878 | * Dirty pagecache page |
6a46079c AK |
879 | * Issues: when the error hit a hole page the error is not properly |
880 | * propagated. | |
881 | */ | |
882 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
883 | { | |
884 | struct address_space *mapping = page_mapping(p); | |
885 | ||
886 | SetPageError(p); | |
887 | /* TBD: print more information about the file. */ | |
888 | if (mapping) { | |
889 | /* | |
890 | * IO error will be reported by write(), fsync(), etc. | |
891 | * who check the mapping. | |
892 | * This way the application knows that something went | |
893 | * wrong with its dirty file data. | |
894 | * | |
895 | * There's one open issue: | |
896 | * | |
897 | * The EIO will be only reported on the next IO | |
898 | * operation and then cleared through the IO map. | |
899 | * Normally Linux has two mechanisms to pass IO error | |
900 | * first through the AS_EIO flag in the address space | |
901 | * and then through the PageError flag in the page. | |
902 | * Since we drop pages on memory failure handling the | |
903 | * only mechanism open to use is through AS_AIO. | |
904 | * | |
905 | * This has the disadvantage that it gets cleared on | |
906 | * the first operation that returns an error, while | |
907 | * the PageError bit is more sticky and only cleared | |
908 | * when the page is reread or dropped. If an | |
909 | * application assumes it will always get error on | |
910 | * fsync, but does other operations on the fd before | |
25985edc | 911 | * and the page is dropped between then the error |
6a46079c AK |
912 | * will not be properly reported. |
913 | * | |
914 | * This can already happen even without hwpoisoned | |
915 | * pages: first on metadata IO errors (which only | |
916 | * report through AS_EIO) or when the page is dropped | |
917 | * at the wrong time. | |
918 | * | |
919 | * So right now we assume that the application DTRT on | |
920 | * the first EIO, but we're not worse than other parts | |
921 | * of the kernel. | |
922 | */ | |
af21bfaf | 923 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
924 | } |
925 | ||
926 | return me_pagecache_clean(p, pfn); | |
927 | } | |
928 | ||
929 | /* | |
930 | * Clean and dirty swap cache. | |
931 | * | |
932 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
933 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
934 | * referenced concurrently by 2 types of PTEs: | |
935 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
936 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
937 | * and then | |
938 | * - clear dirty bit to prevent IO | |
939 | * - remove from LRU | |
940 | * - but keep in the swap cache, so that when we return to it on | |
941 | * a later page fault, we know the application is accessing | |
942 | * corrupted data and shall be killed (we installed simple | |
943 | * interception code in do_swap_page to catch it). | |
944 | * | |
945 | * Clean swap cache pages can be directly isolated. A later page fault will | |
946 | * bring in the known good data from disk. | |
947 | */ | |
948 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
949 | { | |
ea6d0630 NH |
950 | int ret; |
951 | ||
6a46079c AK |
952 | ClearPageDirty(p); |
953 | /* Trigger EIO in shmem: */ | |
954 | ClearPageUptodate(p); | |
955 | ||
ea6d0630 NH |
956 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED; |
957 | unlock_page(p); | |
958 | return ret; | |
6a46079c AK |
959 | } |
960 | ||
961 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
962 | { | |
ea6d0630 NH |
963 | int ret; |
964 | ||
6a46079c | 965 | delete_from_swap_cache(p); |
e43c3afb | 966 | |
ea6d0630 NH |
967 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED; |
968 | unlock_page(p); | |
969 | return ret; | |
6a46079c AK |
970 | } |
971 | ||
972 | /* | |
973 | * Huge pages. Needs work. | |
974 | * Issues: | |
93f70f90 NH |
975 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
976 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
977 | */ |
978 | static int me_huge_page(struct page *p, unsigned long pfn) | |
979 | { | |
a8b2c2ce | 980 | int res; |
93f70f90 | 981 | struct page *hpage = compound_head(p); |
78bb9203 | 982 | struct address_space *mapping; |
2491ffee NH |
983 | |
984 | if (!PageHuge(hpage)) | |
985 | return MF_DELAYED; | |
986 | ||
78bb9203 NH |
987 | mapping = page_mapping(hpage); |
988 | if (mapping) { | |
989 | res = truncate_error_page(hpage, pfn, mapping); | |
ea6d0630 | 990 | unlock_page(hpage); |
78bb9203 | 991 | } else { |
a8b2c2ce | 992 | res = MF_FAILED; |
78bb9203 NH |
993 | unlock_page(hpage); |
994 | /* | |
995 | * migration entry prevents later access on error anonymous | |
996 | * hugepage, so we can free and dissolve it into buddy to | |
997 | * save healthy subpages. | |
998 | */ | |
999 | if (PageAnon(hpage)) | |
1000 | put_page(hpage); | |
510d25c9 | 1001 | if (__page_handle_poison(p)) { |
a8b2c2ce OS |
1002 | page_ref_inc(p); |
1003 | res = MF_RECOVERED; | |
1004 | } | |
93f70f90 | 1005 | } |
78bb9203 NH |
1006 | |
1007 | return res; | |
6a46079c AK |
1008 | } |
1009 | ||
1010 | /* | |
1011 | * Various page states we can handle. | |
1012 | * | |
1013 | * A page state is defined by its current page->flags bits. | |
1014 | * The table matches them in order and calls the right handler. | |
1015 | * | |
1016 | * This is quite tricky because we can access page at any time | |
25985edc | 1017 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
1018 | * |
1019 | * This is not complete. More states could be added. | |
1020 | * For any missing state don't attempt recovery. | |
1021 | */ | |
1022 | ||
1023 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 1024 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
1025 | #define unevict (1UL << PG_unevictable) |
1026 | #define mlock (1UL << PG_mlocked) | |
6a46079c | 1027 | #define lru (1UL << PG_lru) |
6a46079c | 1028 | #define head (1UL << PG_head) |
6a46079c | 1029 | #define slab (1UL << PG_slab) |
6a46079c AK |
1030 | #define reserved (1UL << PG_reserved) |
1031 | ||
1032 | static struct page_state { | |
1033 | unsigned long mask; | |
1034 | unsigned long res; | |
cc637b17 | 1035 | enum mf_action_page_type type; |
ea6d0630 NH |
1036 | |
1037 | /* Callback ->action() has to unlock the relevant page inside it. */ | |
6a46079c AK |
1038 | int (*action)(struct page *p, unsigned long pfn); |
1039 | } error_states[] = { | |
cc637b17 | 1040 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
1041 | /* |
1042 | * free pages are specially detected outside this table: | |
1043 | * PG_buddy pages only make a small fraction of all free pages. | |
1044 | */ | |
6a46079c AK |
1045 | |
1046 | /* | |
1047 | * Could in theory check if slab page is free or if we can drop | |
1048 | * currently unused objects without touching them. But just | |
1049 | * treat it as standard kernel for now. | |
1050 | */ | |
cc637b17 | 1051 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 1052 | |
cc637b17 | 1053 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 1054 | |
cc637b17 XX |
1055 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
1056 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 1057 | |
cc637b17 XX |
1058 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
1059 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 1060 | |
cc637b17 XX |
1061 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
1062 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 1063 | |
cc637b17 XX |
1064 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
1065 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
1066 | |
1067 | /* | |
1068 | * Catchall entry: must be at end. | |
1069 | */ | |
cc637b17 | 1070 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
1071 | }; |
1072 | ||
2326c467 AK |
1073 | #undef dirty |
1074 | #undef sc | |
1075 | #undef unevict | |
1076 | #undef mlock | |
2326c467 | 1077 | #undef lru |
2326c467 | 1078 | #undef head |
2326c467 AK |
1079 | #undef slab |
1080 | #undef reserved | |
1081 | ||
ff604cf6 NH |
1082 | /* |
1083 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
1084 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
1085 | */ | |
cc3e2af4 XX |
1086 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
1087 | enum mf_result result) | |
6a46079c | 1088 | { |
97f0b134 XX |
1089 | trace_memory_failure_event(pfn, type, result); |
1090 | ||
495367c0 | 1091 | pr_err("Memory failure: %#lx: recovery action for %s: %s\n", |
64d37a2b | 1092 | pfn, action_page_types[type], action_name[result]); |
6a46079c AK |
1093 | } |
1094 | ||
1095 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 1096 | unsigned long pfn) |
6a46079c AK |
1097 | { |
1098 | int result; | |
7456b040 | 1099 | int count; |
6a46079c | 1100 | |
ea6d0630 | 1101 | /* page p should be unlocked after returning from ps->action(). */ |
6a46079c | 1102 | result = ps->action(p, pfn); |
7456b040 | 1103 | |
bd1ce5f9 | 1104 | count = page_count(p) - 1; |
cc637b17 | 1105 | if (ps->action == me_swapcache_dirty && result == MF_DELAYED) |
138ce286 | 1106 | count--; |
78bb9203 | 1107 | if (count > 0) { |
495367c0 | 1108 | pr_err("Memory failure: %#lx: %s still referenced by %d users\n", |
64d37a2b | 1109 | pfn, action_page_types[ps->type], count); |
cc637b17 | 1110 | result = MF_FAILED; |
138ce286 | 1111 | } |
64d37a2b | 1112 | action_result(pfn, ps->type, result); |
6a46079c AK |
1113 | |
1114 | /* Could do more checks here if page looks ok */ | |
1115 | /* | |
1116 | * Could adjust zone counters here to correct for the missing page. | |
1117 | */ | |
1118 | ||
cc637b17 | 1119 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
1120 | } |
1121 | ||
25182f05 NH |
1122 | /* |
1123 | * Return true if a page type of a given page is supported by hwpoison | |
1124 | * mechanism (while handling could fail), otherwise false. This function | |
1125 | * does not return true for hugetlb or device memory pages, so it's assumed | |
1126 | * to be called only in the context where we never have such pages. | |
1127 | */ | |
1128 | static inline bool HWPoisonHandlable(struct page *page) | |
1129 | { | |
1130 | return PageLRU(page) || __PageMovable(page); | |
1131 | } | |
1132 | ||
17e395b6 | 1133 | static int __get_hwpoison_page(struct page *page) |
ead07f6a NH |
1134 | { |
1135 | struct page *head = compound_head(page); | |
25182f05 NH |
1136 | int ret = 0; |
1137 | bool hugetlb = false; | |
1138 | ||
1139 | ret = get_hwpoison_huge_page(head, &hugetlb); | |
1140 | if (hugetlb) | |
1141 | return ret; | |
1142 | ||
1143 | /* | |
1144 | * This check prevents from calling get_hwpoison_unless_zero() | |
1145 | * for any unsupported type of page in order to reduce the risk of | |
1146 | * unexpected races caused by taking a page refcount. | |
1147 | */ | |
1148 | if (!HWPoisonHandlable(head)) | |
1149 | return 0; | |
ead07f6a | 1150 | |
25182f05 | 1151 | if (PageTransHuge(head)) { |
98ed2b00 NH |
1152 | /* |
1153 | * Non anonymous thp exists only in allocation/free time. We | |
1154 | * can't handle such a case correctly, so let's give it up. | |
1155 | * This should be better than triggering BUG_ON when kernel | |
1156 | * tries to touch the "partially handled" page. | |
1157 | */ | |
1158 | if (!PageAnon(head)) { | |
495367c0 | 1159 | pr_err("Memory failure: %#lx: non anonymous thp\n", |
98ed2b00 NH |
1160 | page_to_pfn(page)); |
1161 | return 0; | |
1162 | } | |
ead07f6a NH |
1163 | } |
1164 | ||
c2e7e00b KK |
1165 | if (get_page_unless_zero(head)) { |
1166 | if (head == compound_head(page)) | |
1167 | return 1; | |
1168 | ||
495367c0 CY |
1169 | pr_info("Memory failure: %#lx cannot catch tail\n", |
1170 | page_to_pfn(page)); | |
c2e7e00b KK |
1171 | put_page(head); |
1172 | } | |
1173 | ||
1174 | return 0; | |
ead07f6a | 1175 | } |
ead07f6a | 1176 | |
2f714160 | 1177 | static int get_any_page(struct page *p, unsigned long flags) |
17e395b6 | 1178 | { |
2f714160 OS |
1179 | int ret = 0, pass = 0; |
1180 | bool count_increased = false; | |
17e395b6 | 1181 | |
2f714160 OS |
1182 | if (flags & MF_COUNT_INCREASED) |
1183 | count_increased = true; | |
1184 | ||
1185 | try_again: | |
0ed950d1 NH |
1186 | if (!count_increased) { |
1187 | ret = __get_hwpoison_page(p); | |
1188 | if (!ret) { | |
1189 | if (page_count(p)) { | |
1190 | /* We raced with an allocation, retry. */ | |
1191 | if (pass++ < 3) | |
1192 | goto try_again; | |
1193 | ret = -EBUSY; | |
1194 | } else if (!PageHuge(p) && !is_free_buddy_page(p)) { | |
1195 | /* We raced with put_page, retry. */ | |
1196 | if (pass++ < 3) | |
1197 | goto try_again; | |
1198 | ret = -EIO; | |
1199 | } | |
1200 | goto out; | |
1201 | } else if (ret == -EBUSY) { | |
1202 | /* We raced with freeing huge page to buddy, retry. */ | |
2f714160 OS |
1203 | if (pass++ < 3) |
1204 | goto try_again; | |
0ed950d1 | 1205 | goto out; |
2f714160 | 1206 | } |
0ed950d1 NH |
1207 | } |
1208 | ||
1209 | if (PageHuge(p) || HWPoisonHandlable(p)) { | |
1210 | ret = 1; | |
2f714160 | 1211 | } else { |
0ed950d1 NH |
1212 | /* |
1213 | * A page we cannot handle. Check whether we can turn | |
1214 | * it into something we can handle. | |
1215 | */ | |
1216 | if (pass++ < 3) { | |
2f714160 | 1217 | put_page(p); |
0ed950d1 NH |
1218 | shake_page(p, 1); |
1219 | count_increased = false; | |
1220 | goto try_again; | |
2f714160 | 1221 | } |
0ed950d1 NH |
1222 | put_page(p); |
1223 | ret = -EIO; | |
17e395b6 | 1224 | } |
0ed950d1 | 1225 | out: |
17e395b6 OS |
1226 | return ret; |
1227 | } | |
1228 | ||
0ed950d1 NH |
1229 | /** |
1230 | * get_hwpoison_page() - Get refcount for memory error handling | |
1231 | * @p: Raw error page (hit by memory error) | |
1232 | * @flags: Flags controlling behavior of error handling | |
1233 | * | |
1234 | * get_hwpoison_page() takes a page refcount of an error page to handle memory | |
1235 | * error on it, after checking that the error page is in a well-defined state | |
1236 | * (defined as a page-type we can successfully handle the memor error on it, | |
1237 | * such as LRU page and hugetlb page). | |
1238 | * | |
1239 | * Memory error handling could be triggered at any time on any type of page, | |
1240 | * so it's prone to race with typical memory management lifecycle (like | |
1241 | * allocation and free). So to avoid such races, get_hwpoison_page() takes | |
1242 | * extra care for the error page's state (as done in __get_hwpoison_page()), | |
1243 | * and has some retry logic in get_any_page(). | |
1244 | * | |
1245 | * Return: 0 on failure, | |
1246 | * 1 on success for in-use pages in a well-defined state, | |
1247 | * -EIO for pages on which we can not handle memory errors, | |
1248 | * -EBUSY when get_hwpoison_page() has raced with page lifecycle | |
1249 | * operations like allocation and free. | |
1250 | */ | |
1251 | static int get_hwpoison_page(struct page *p, unsigned long flags) | |
2f714160 OS |
1252 | { |
1253 | int ret; | |
1254 | ||
1255 | zone_pcp_disable(page_zone(p)); | |
0ed950d1 | 1256 | ret = get_any_page(p, flags); |
2f714160 OS |
1257 | zone_pcp_enable(page_zone(p)); |
1258 | ||
1259 | return ret; | |
1260 | } | |
1261 | ||
6a46079c AK |
1262 | /* |
1263 | * Do all that is necessary to remove user space mappings. Unmap | |
1264 | * the pages and send SIGBUS to the processes if the data was dirty. | |
1265 | */ | |
666e5a40 | 1266 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
83b57531 | 1267 | int flags, struct page **hpagep) |
6a46079c | 1268 | { |
36af6737 | 1269 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC; |
6a46079c AK |
1270 | struct address_space *mapping; |
1271 | LIST_HEAD(tokill); | |
1fb08ac6 | 1272 | bool unmap_success; |
6751ed65 | 1273 | int kill = 1, forcekill; |
54b9dd14 | 1274 | struct page *hpage = *hpagep; |
286c469a | 1275 | bool mlocked = PageMlocked(hpage); |
6a46079c | 1276 | |
93a9eb39 NH |
1277 | /* |
1278 | * Here we are interested only in user-mapped pages, so skip any | |
1279 | * other types of pages. | |
1280 | */ | |
1281 | if (PageReserved(p) || PageSlab(p)) | |
666e5a40 | 1282 | return true; |
93a9eb39 | 1283 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 1284 | return true; |
6a46079c | 1285 | |
6a46079c AK |
1286 | /* |
1287 | * This check implies we don't kill processes if their pages | |
1288 | * are in the swap cache early. Those are always late kills. | |
1289 | */ | |
7af446a8 | 1290 | if (!page_mapped(hpage)) |
666e5a40 | 1291 | return true; |
1668bfd5 | 1292 | |
52089b14 | 1293 | if (PageKsm(p)) { |
495367c0 | 1294 | pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn); |
666e5a40 | 1295 | return false; |
52089b14 | 1296 | } |
6a46079c AK |
1297 | |
1298 | if (PageSwapCache(p)) { | |
495367c0 CY |
1299 | pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n", |
1300 | pfn); | |
6a46079c AK |
1301 | ttu |= TTU_IGNORE_HWPOISON; |
1302 | } | |
1303 | ||
1304 | /* | |
1305 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1306 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1307 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1308 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1309 | */ |
7af446a8 | 1310 | mapping = page_mapping(hpage); |
6751ed65 | 1311 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
f56753ac | 1312 | mapping_can_writeback(mapping)) { |
7af446a8 NH |
1313 | if (page_mkclean(hpage)) { |
1314 | SetPageDirty(hpage); | |
6a46079c AK |
1315 | } else { |
1316 | kill = 0; | |
1317 | ttu |= TTU_IGNORE_HWPOISON; | |
495367c0 | 1318 | pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1319 | pfn); |
1320 | } | |
1321 | } | |
1322 | ||
1323 | /* | |
1324 | * First collect all the processes that have the page | |
1325 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1326 | * because ttu takes the rmap data structures down. | |
1327 | * | |
1328 | * Error handling: We ignore errors here because | |
1329 | * there's nothing that can be done. | |
1330 | */ | |
1331 | if (kill) | |
415c64c1 | 1332 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1333 | |
c0d0381a | 1334 | if (!PageHuge(hpage)) { |
1fb08ac6 | 1335 | try_to_unmap(hpage, ttu); |
c0d0381a | 1336 | } else { |
336bf30e MK |
1337 | if (!PageAnon(hpage)) { |
1338 | /* | |
1339 | * For hugetlb pages in shared mappings, try_to_unmap | |
1340 | * could potentially call huge_pmd_unshare. Because of | |
1341 | * this, take semaphore in write mode here and set | |
1342 | * TTU_RMAP_LOCKED to indicate we have taken the lock | |
041711ce | 1343 | * at this higher level. |
336bf30e MK |
1344 | */ |
1345 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1346 | if (mapping) { | |
1fb08ac6 | 1347 | try_to_unmap(hpage, ttu|TTU_RMAP_LOCKED); |
336bf30e | 1348 | i_mmap_unlock_write(mapping); |
1fb08ac6 | 1349 | } else |
336bf30e | 1350 | pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn); |
c0d0381a | 1351 | } else { |
1fb08ac6 | 1352 | try_to_unmap(hpage, ttu); |
c0d0381a MK |
1353 | } |
1354 | } | |
1fb08ac6 YS |
1355 | |
1356 | unmap_success = !page_mapped(hpage); | |
666e5a40 | 1357 | if (!unmap_success) |
495367c0 | 1358 | pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1359 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1360 | |
286c469a NH |
1361 | /* |
1362 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1363 | * shake_page() again to ensure that it's flushed. | |
1364 | */ | |
1365 | if (mlocked) | |
1366 | shake_page(hpage, 0); | |
1367 | ||
6a46079c AK |
1368 | /* |
1369 | * Now that the dirty bit has been propagated to the | |
1370 | * struct page and all unmaps done we can decide if | |
1371 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1372 | * was dirty or the process is not restartable, |
1373 | * otherwise the tokill list is merely | |
6a46079c AK |
1374 | * freed. When there was a problem unmapping earlier |
1375 | * use a more force-full uncatchable kill to prevent | |
1376 | * any accesses to the poisoned memory. | |
1377 | */ | |
415c64c1 | 1378 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
ae1139ec | 1379 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1380 | |
666e5a40 | 1381 | return unmap_success; |
6a46079c AK |
1382 | } |
1383 | ||
0348d2eb NH |
1384 | static int identify_page_state(unsigned long pfn, struct page *p, |
1385 | unsigned long page_flags) | |
761ad8d7 NH |
1386 | { |
1387 | struct page_state *ps; | |
0348d2eb NH |
1388 | |
1389 | /* | |
1390 | * The first check uses the current page flags which may not have any | |
1391 | * relevant information. The second check with the saved page flags is | |
1392 | * carried out only if the first check can't determine the page status. | |
1393 | */ | |
1394 | for (ps = error_states;; ps++) | |
1395 | if ((p->flags & ps->mask) == ps->res) | |
1396 | break; | |
1397 | ||
1398 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1399 | ||
1400 | if (!ps->mask) | |
1401 | for (ps = error_states;; ps++) | |
1402 | if ((page_flags & ps->mask) == ps->res) | |
1403 | break; | |
1404 | return page_action(ps, p, pfn); | |
1405 | } | |
1406 | ||
694bf0b0 OS |
1407 | static int try_to_split_thp_page(struct page *page, const char *msg) |
1408 | { | |
1409 | lock_page(page); | |
1410 | if (!PageAnon(page) || unlikely(split_huge_page(page))) { | |
1411 | unsigned long pfn = page_to_pfn(page); | |
1412 | ||
1413 | unlock_page(page); | |
1414 | if (!PageAnon(page)) | |
1415 | pr_info("%s: %#lx: non anonymous thp\n", msg, pfn); | |
1416 | else | |
1417 | pr_info("%s: %#lx: thp split failed\n", msg, pfn); | |
1418 | put_page(page); | |
1419 | return -EBUSY; | |
1420 | } | |
1421 | unlock_page(page); | |
1422 | ||
1423 | return 0; | |
1424 | } | |
1425 | ||
83b57531 | 1426 | static int memory_failure_hugetlb(unsigned long pfn, int flags) |
0348d2eb | 1427 | { |
761ad8d7 NH |
1428 | struct page *p = pfn_to_page(pfn); |
1429 | struct page *head = compound_head(p); | |
1430 | int res; | |
1431 | unsigned long page_flags; | |
1432 | ||
1433 | if (TestSetPageHWPoison(head)) { | |
1434 | pr_err("Memory failure: %#lx: already hardware poisoned\n", | |
1435 | pfn); | |
a3f5d80e NH |
1436 | res = -EHWPOISON; |
1437 | if (flags & MF_ACTION_REQUIRED) | |
1438 | res = kill_accessing_process(current, page_to_pfn(head), flags); | |
1439 | return res; | |
761ad8d7 NH |
1440 | } |
1441 | ||
1442 | num_poisoned_pages_inc(); | |
1443 | ||
0ed950d1 NH |
1444 | if (!(flags & MF_COUNT_INCREASED)) { |
1445 | res = get_hwpoison_page(p, flags); | |
1446 | if (!res) { | |
1447 | /* | |
1448 | * Check "filter hit" and "race with other subpage." | |
1449 | */ | |
1450 | lock_page(head); | |
1451 | if (PageHWPoison(head)) { | |
1452 | if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1453 | || (p != head && TestSetPageHWPoison(head))) { | |
1454 | num_poisoned_pages_dec(); | |
1455 | unlock_page(head); | |
1456 | return 0; | |
1457 | } | |
761ad8d7 | 1458 | } |
0ed950d1 NH |
1459 | unlock_page(head); |
1460 | res = MF_FAILED; | |
510d25c9 | 1461 | if (__page_handle_poison(p)) { |
0ed950d1 NH |
1462 | page_ref_inc(p); |
1463 | res = MF_RECOVERED; | |
1464 | } | |
1465 | action_result(pfn, MF_MSG_FREE_HUGE, res); | |
1466 | return res == MF_RECOVERED ? 0 : -EBUSY; | |
1467 | } else if (res < 0) { | |
1468 | action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); | |
1469 | return -EBUSY; | |
761ad8d7 | 1470 | } |
761ad8d7 NH |
1471 | } |
1472 | ||
1473 | lock_page(head); | |
1474 | page_flags = head->flags; | |
1475 | ||
1476 | if (!PageHWPoison(head)) { | |
1477 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); | |
1478 | num_poisoned_pages_dec(); | |
1479 | unlock_page(head); | |
dd6e2402 | 1480 | put_page(head); |
761ad8d7 NH |
1481 | return 0; |
1482 | } | |
1483 | ||
31286a84 NH |
1484 | /* |
1485 | * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so | |
1486 | * simply disable it. In order to make it work properly, we need | |
1487 | * make sure that: | |
1488 | * - conversion of a pud that maps an error hugetlb into hwpoison | |
1489 | * entry properly works, and | |
1490 | * - other mm code walking over page table is aware of pud-aligned | |
1491 | * hwpoison entries. | |
1492 | */ | |
1493 | if (huge_page_size(page_hstate(head)) > PMD_SIZE) { | |
1494 | action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED); | |
1495 | res = -EBUSY; | |
1496 | goto out; | |
1497 | } | |
1498 | ||
83b57531 | 1499 | if (!hwpoison_user_mappings(p, pfn, flags, &head)) { |
761ad8d7 NH |
1500 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1501 | res = -EBUSY; | |
1502 | goto out; | |
1503 | } | |
1504 | ||
ea6d0630 | 1505 | return identify_page_state(pfn, p, page_flags); |
761ad8d7 NH |
1506 | out: |
1507 | unlock_page(head); | |
1508 | return res; | |
1509 | } | |
1510 | ||
6100e34b DW |
1511 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
1512 | struct dev_pagemap *pgmap) | |
1513 | { | |
1514 | struct page *page = pfn_to_page(pfn); | |
1515 | const bool unmap_success = true; | |
1516 | unsigned long size = 0; | |
1517 | struct to_kill *tk; | |
1518 | LIST_HEAD(tokill); | |
1519 | int rc = -EBUSY; | |
1520 | loff_t start; | |
27359fd6 | 1521 | dax_entry_t cookie; |
6100e34b | 1522 | |
1e8aaedb OS |
1523 | if (flags & MF_COUNT_INCREASED) |
1524 | /* | |
1525 | * Drop the extra refcount in case we come from madvise(). | |
1526 | */ | |
1527 | put_page(page); | |
1528 | ||
34dc45be DW |
1529 | /* device metadata space is not recoverable */ |
1530 | if (!pgmap_pfn_valid(pgmap, pfn)) { | |
1531 | rc = -ENXIO; | |
1532 | goto out; | |
1533 | } | |
1534 | ||
6100e34b DW |
1535 | /* |
1536 | * Prevent the inode from being freed while we are interrogating | |
1537 | * the address_space, typically this would be handled by | |
1538 | * lock_page(), but dax pages do not use the page lock. This | |
1539 | * also prevents changes to the mapping of this pfn until | |
1540 | * poison signaling is complete. | |
1541 | */ | |
27359fd6 MW |
1542 | cookie = dax_lock_page(page); |
1543 | if (!cookie) | |
6100e34b DW |
1544 | goto out; |
1545 | ||
1546 | if (hwpoison_filter(page)) { | |
1547 | rc = 0; | |
1548 | goto unlock; | |
1549 | } | |
1550 | ||
25b2995a | 1551 | if (pgmap->type == MEMORY_DEVICE_PRIVATE) { |
6100e34b DW |
1552 | /* |
1553 | * TODO: Handle HMM pages which may need coordination | |
1554 | * with device-side memory. | |
1555 | */ | |
1556 | goto unlock; | |
6100e34b DW |
1557 | } |
1558 | ||
1559 | /* | |
1560 | * Use this flag as an indication that the dax page has been | |
1561 | * remapped UC to prevent speculative consumption of poison. | |
1562 | */ | |
1563 | SetPageHWPoison(page); | |
1564 | ||
1565 | /* | |
1566 | * Unlike System-RAM there is no possibility to swap in a | |
1567 | * different physical page at a given virtual address, so all | |
1568 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1569 | * SIGBUS (i.e. MF_MUST_KILL) | |
1570 | */ | |
1571 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1572 | collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED); | |
1573 | ||
1574 | list_for_each_entry(tk, &tokill, nd) | |
1575 | if (tk->size_shift) | |
1576 | size = max(size, 1UL << tk->size_shift); | |
1577 | if (size) { | |
1578 | /* | |
1579 | * Unmap the largest mapping to avoid breaking up | |
1580 | * device-dax mappings which are constant size. The | |
1581 | * actual size of the mapping being torn down is | |
1582 | * communicated in siginfo, see kill_proc() | |
1583 | */ | |
1584 | start = (page->index << PAGE_SHIFT) & ~(size - 1); | |
4d75136b | 1585 | unmap_mapping_range(page->mapping, start, size, 0); |
6100e34b DW |
1586 | } |
1587 | kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags); | |
1588 | rc = 0; | |
1589 | unlock: | |
27359fd6 | 1590 | dax_unlock_page(page, cookie); |
6100e34b DW |
1591 | out: |
1592 | /* drop pgmap ref acquired in caller */ | |
1593 | put_dev_pagemap(pgmap); | |
1594 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
1595 | return rc; | |
1596 | } | |
1597 | ||
cd42f4a3 TL |
1598 | /** |
1599 | * memory_failure - Handle memory failure of a page. | |
1600 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
1601 | * @flags: fine tune action taken |
1602 | * | |
1603 | * This function is called by the low level machine check code | |
1604 | * of an architecture when it detects hardware memory corruption | |
1605 | * of a page. It tries its best to recover, which includes | |
1606 | * dropping pages, killing processes etc. | |
1607 | * | |
1608 | * The function is primarily of use for corruptions that | |
1609 | * happen outside the current execution context (e.g. when | |
1610 | * detected by a background scrubber) | |
1611 | * | |
1612 | * Must run in process context (e.g. a work queue) with interrupts | |
1613 | * enabled and no spinlocks hold. | |
1614 | */ | |
83b57531 | 1615 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 1616 | { |
6a46079c | 1617 | struct page *p; |
7af446a8 | 1618 | struct page *hpage; |
415c64c1 | 1619 | struct page *orig_head; |
6100e34b | 1620 | struct dev_pagemap *pgmap; |
171936dd | 1621 | int res = 0; |
524fca1e | 1622 | unsigned long page_flags; |
a8b2c2ce | 1623 | bool retry = true; |
171936dd | 1624 | static DEFINE_MUTEX(mf_mutex); |
6a46079c AK |
1625 | |
1626 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 1627 | panic("Memory failure on page %lx", pfn); |
6a46079c | 1628 | |
96c804a6 DH |
1629 | p = pfn_to_online_page(pfn); |
1630 | if (!p) { | |
1631 | if (pfn_valid(pfn)) { | |
1632 | pgmap = get_dev_pagemap(pfn, NULL); | |
1633 | if (pgmap) | |
1634 | return memory_failure_dev_pagemap(pfn, flags, | |
1635 | pgmap); | |
1636 | } | |
495367c0 CY |
1637 | pr_err("Memory failure: %#lx: memory outside kernel control\n", |
1638 | pfn); | |
a7560fc8 | 1639 | return -ENXIO; |
6a46079c AK |
1640 | } |
1641 | ||
171936dd TL |
1642 | mutex_lock(&mf_mutex); |
1643 | ||
a8b2c2ce | 1644 | try_again: |
171936dd TL |
1645 | if (PageHuge(p)) { |
1646 | res = memory_failure_hugetlb(pfn, flags); | |
1647 | goto unlock_mutex; | |
1648 | } | |
1649 | ||
6a46079c | 1650 | if (TestSetPageHWPoison(p)) { |
495367c0 CY |
1651 | pr_err("Memory failure: %#lx: already hardware poisoned\n", |
1652 | pfn); | |
47af12ba | 1653 | res = -EHWPOISON; |
a3f5d80e NH |
1654 | if (flags & MF_ACTION_REQUIRED) |
1655 | res = kill_accessing_process(current, pfn, flags); | |
171936dd | 1656 | goto unlock_mutex; |
6a46079c AK |
1657 | } |
1658 | ||
761ad8d7 | 1659 | orig_head = hpage = compound_head(p); |
b37ff71c | 1660 | num_poisoned_pages_inc(); |
6a46079c AK |
1661 | |
1662 | /* | |
1663 | * We need/can do nothing about count=0 pages. | |
1664 | * 1) it's a free page, and therefore in safe hand: | |
1665 | * prep_new_page() will be the gate keeper. | |
761ad8d7 | 1666 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
1667 | * Implies some kernel user: cannot stop them from |
1668 | * R/W the page; let's pray that the page has been | |
1669 | * used and will be freed some time later. | |
1670 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 1671 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 1672 | */ |
0ed950d1 NH |
1673 | if (!(flags & MF_COUNT_INCREASED)) { |
1674 | res = get_hwpoison_page(p, flags); | |
1675 | if (!res) { | |
1676 | if (is_free_buddy_page(p)) { | |
1677 | if (take_page_off_buddy(p)) { | |
1678 | page_ref_inc(p); | |
1679 | res = MF_RECOVERED; | |
1680 | } else { | |
1681 | /* We lost the race, try again */ | |
1682 | if (retry) { | |
1683 | ClearPageHWPoison(p); | |
1684 | num_poisoned_pages_dec(); | |
1685 | retry = false; | |
1686 | goto try_again; | |
1687 | } | |
1688 | res = MF_FAILED; | |
a8b2c2ce | 1689 | } |
0ed950d1 NH |
1690 | action_result(pfn, MF_MSG_BUDDY, res); |
1691 | res = res == MF_RECOVERED ? 0 : -EBUSY; | |
1692 | } else { | |
1693 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); | |
1694 | res = -EBUSY; | |
a8b2c2ce | 1695 | } |
0ed950d1 NH |
1696 | goto unlock_mutex; |
1697 | } else if (res < 0) { | |
1698 | action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); | |
171936dd | 1699 | res = -EBUSY; |
0ed950d1 | 1700 | goto unlock_mutex; |
8d22ba1b | 1701 | } |
6a46079c AK |
1702 | } |
1703 | ||
761ad8d7 | 1704 | if (PageTransHuge(hpage)) { |
5d1fd5dc NH |
1705 | if (try_to_split_thp_page(p, "Memory Failure") < 0) { |
1706 | action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED); | |
171936dd TL |
1707 | res = -EBUSY; |
1708 | goto unlock_mutex; | |
5d1fd5dc | 1709 | } |
415c64c1 | 1710 | VM_BUG_ON_PAGE(!page_count(p), p); |
415c64c1 NH |
1711 | } |
1712 | ||
e43c3afb WF |
1713 | /* |
1714 | * We ignore non-LRU pages for good reasons. | |
1715 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 1716 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
1717 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
1718 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1719 | * walked by the page reclaim code, however that's not a big loss. | |
1720 | */ | |
8bcb74de | 1721 | shake_page(p, 0); |
e43c3afb | 1722 | |
761ad8d7 | 1723 | lock_page(p); |
847ce401 | 1724 | |
f37d4298 AK |
1725 | /* |
1726 | * The page could have changed compound pages during the locking. | |
1727 | * If this happens just bail out. | |
1728 | */ | |
415c64c1 | 1729 | if (PageCompound(p) && compound_head(p) != orig_head) { |
cc637b17 | 1730 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 | 1731 | res = -EBUSY; |
171936dd | 1732 | goto unlock_page; |
f37d4298 AK |
1733 | } |
1734 | ||
524fca1e NH |
1735 | /* |
1736 | * We use page flags to determine what action should be taken, but | |
1737 | * the flags can be modified by the error containment action. One | |
1738 | * example is an mlocked page, where PG_mlocked is cleared by | |
1739 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1740 | * correctly, we save a copy of the page flags at this time. | |
1741 | */ | |
7d9d46ac | 1742 | page_flags = p->flags; |
524fca1e | 1743 | |
847ce401 WF |
1744 | /* |
1745 | * unpoison always clear PG_hwpoison inside page lock | |
1746 | */ | |
1747 | if (!PageHWPoison(p)) { | |
495367c0 | 1748 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); |
b37ff71c | 1749 | num_poisoned_pages_dec(); |
761ad8d7 | 1750 | unlock_page(p); |
dd6e2402 | 1751 | put_page(p); |
171936dd | 1752 | goto unlock_mutex; |
847ce401 | 1753 | } |
7c116f2b WF |
1754 | if (hwpoison_filter(p)) { |
1755 | if (TestClearPageHWPoison(p)) | |
b37ff71c | 1756 | num_poisoned_pages_dec(); |
761ad8d7 | 1757 | unlock_page(p); |
dd6e2402 | 1758 | put_page(p); |
171936dd | 1759 | goto unlock_mutex; |
7c116f2b | 1760 | } |
847ce401 | 1761 | |
e8675d29 | 1762 | /* |
1763 | * __munlock_pagevec may clear a writeback page's LRU flag without | |
1764 | * page_lock. We need wait writeback completion for this page or it | |
1765 | * may trigger vfs BUG while evict inode. | |
1766 | */ | |
1767 | if (!PageTransTail(p) && !PageLRU(p) && !PageWriteback(p)) | |
0bc1f8b0 CY |
1768 | goto identify_page_state; |
1769 | ||
6edd6cc6 NH |
1770 | /* |
1771 | * It's very difficult to mess with pages currently under IO | |
1772 | * and in many cases impossible, so we just avoid it here. | |
1773 | */ | |
6a46079c AK |
1774 | wait_on_page_writeback(p); |
1775 | ||
1776 | /* | |
1777 | * Now take care of user space mappings. | |
e64a782f | 1778 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
6a46079c | 1779 | */ |
1b473bec | 1780 | if (!hwpoison_user_mappings(p, pfn, flags, &p)) { |
cc637b17 | 1781 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 | 1782 | res = -EBUSY; |
171936dd | 1783 | goto unlock_page; |
1668bfd5 | 1784 | } |
6a46079c AK |
1785 | |
1786 | /* | |
1787 | * Torn down by someone else? | |
1788 | */ | |
dc2a1cbf | 1789 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 1790 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 1791 | res = -EBUSY; |
171936dd | 1792 | goto unlock_page; |
6a46079c AK |
1793 | } |
1794 | ||
0bc1f8b0 | 1795 | identify_page_state: |
0348d2eb | 1796 | res = identify_page_state(pfn, p, page_flags); |
ea6d0630 NH |
1797 | mutex_unlock(&mf_mutex); |
1798 | return res; | |
171936dd | 1799 | unlock_page: |
761ad8d7 | 1800 | unlock_page(p); |
171936dd TL |
1801 | unlock_mutex: |
1802 | mutex_unlock(&mf_mutex); | |
6a46079c AK |
1803 | return res; |
1804 | } | |
cd42f4a3 | 1805 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1806 | |
ea8f5fb8 HY |
1807 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1808 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1809 | ||
1810 | struct memory_failure_entry { | |
1811 | unsigned long pfn; | |
ea8f5fb8 HY |
1812 | int flags; |
1813 | }; | |
1814 | ||
1815 | struct memory_failure_cpu { | |
1816 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1817 | MEMORY_FAILURE_FIFO_SIZE); | |
1818 | spinlock_t lock; | |
1819 | struct work_struct work; | |
1820 | }; | |
1821 | ||
1822 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1823 | ||
1824 | /** | |
1825 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1826 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
1827 | * @flags: Flags for memory failure handling |
1828 | * | |
1829 | * This function is called by the low level hardware error handler | |
1830 | * when it detects hardware memory corruption of a page. It schedules | |
1831 | * the recovering of error page, including dropping pages, killing | |
1832 | * processes etc. | |
1833 | * | |
1834 | * The function is primarily of use for corruptions that | |
1835 | * happen outside the current execution context (e.g. when | |
1836 | * detected by a background scrubber) | |
1837 | * | |
1838 | * Can run in IRQ context. | |
1839 | */ | |
83b57531 | 1840 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
1841 | { |
1842 | struct memory_failure_cpu *mf_cpu; | |
1843 | unsigned long proc_flags; | |
1844 | struct memory_failure_entry entry = { | |
1845 | .pfn = pfn, | |
ea8f5fb8 HY |
1846 | .flags = flags, |
1847 | }; | |
1848 | ||
1849 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1850 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 1851 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
1852 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
1853 | else | |
8e33a52f | 1854 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
1855 | pfn); |
1856 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1857 | put_cpu_var(memory_failure_cpu); | |
1858 | } | |
1859 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1860 | ||
1861 | static void memory_failure_work_func(struct work_struct *work) | |
1862 | { | |
1863 | struct memory_failure_cpu *mf_cpu; | |
1864 | struct memory_failure_entry entry = { 0, }; | |
1865 | unsigned long proc_flags; | |
1866 | int gotten; | |
1867 | ||
06202231 | 1868 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
ea8f5fb8 HY |
1869 | for (;;) { |
1870 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1871 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1872 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1873 | if (!gotten) | |
1874 | break; | |
cf870c70 | 1875 | if (entry.flags & MF_SOFT_OFFLINE) |
feec24a6 | 1876 | soft_offline_page(entry.pfn, entry.flags); |
cf870c70 | 1877 | else |
83b57531 | 1878 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
1879 | } |
1880 | } | |
1881 | ||
06202231 JM |
1882 | /* |
1883 | * Process memory_failure work queued on the specified CPU. | |
1884 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. | |
1885 | */ | |
1886 | void memory_failure_queue_kick(int cpu) | |
1887 | { | |
1888 | struct memory_failure_cpu *mf_cpu; | |
1889 | ||
1890 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1891 | cancel_work_sync(&mf_cpu->work); | |
1892 | memory_failure_work_func(&mf_cpu->work); | |
1893 | } | |
1894 | ||
ea8f5fb8 HY |
1895 | static int __init memory_failure_init(void) |
1896 | { | |
1897 | struct memory_failure_cpu *mf_cpu; | |
1898 | int cpu; | |
1899 | ||
1900 | for_each_possible_cpu(cpu) { | |
1901 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1902 | spin_lock_init(&mf_cpu->lock); | |
1903 | INIT_KFIFO(mf_cpu->fifo); | |
1904 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1905 | } | |
1906 | ||
1907 | return 0; | |
1908 | } | |
1909 | core_initcall(memory_failure_init); | |
1910 | ||
a5f65109 NH |
1911 | #define unpoison_pr_info(fmt, pfn, rs) \ |
1912 | ({ \ | |
1913 | if (__ratelimit(rs)) \ | |
1914 | pr_info(fmt, pfn); \ | |
1915 | }) | |
1916 | ||
847ce401 WF |
1917 | /** |
1918 | * unpoison_memory - Unpoison a previously poisoned page | |
1919 | * @pfn: Page number of the to be unpoisoned page | |
1920 | * | |
1921 | * Software-unpoison a page that has been poisoned by | |
1922 | * memory_failure() earlier. | |
1923 | * | |
1924 | * This is only done on the software-level, so it only works | |
1925 | * for linux injected failures, not real hardware failures | |
1926 | * | |
1927 | * Returns 0 for success, otherwise -errno. | |
1928 | */ | |
1929 | int unpoison_memory(unsigned long pfn) | |
1930 | { | |
1931 | struct page *page; | |
1932 | struct page *p; | |
1933 | int freeit = 0; | |
2f714160 | 1934 | unsigned long flags = 0; |
a5f65109 NH |
1935 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
1936 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
1937 | |
1938 | if (!pfn_valid(pfn)) | |
1939 | return -ENXIO; | |
1940 | ||
1941 | p = pfn_to_page(pfn); | |
1942 | page = compound_head(p); | |
1943 | ||
1944 | if (!PageHWPoison(p)) { | |
495367c0 | 1945 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 1946 | pfn, &unpoison_rs); |
847ce401 WF |
1947 | return 0; |
1948 | } | |
1949 | ||
230ac719 | 1950 | if (page_count(page) > 1) { |
495367c0 | 1951 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 1952 | pfn, &unpoison_rs); |
230ac719 NH |
1953 | return 0; |
1954 | } | |
1955 | ||
1956 | if (page_mapped(page)) { | |
495367c0 | 1957 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 1958 | pfn, &unpoison_rs); |
230ac719 NH |
1959 | return 0; |
1960 | } | |
1961 | ||
1962 | if (page_mapping(page)) { | |
495367c0 | 1963 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 1964 | pfn, &unpoison_rs); |
230ac719 NH |
1965 | return 0; |
1966 | } | |
1967 | ||
0cea3fdc WL |
1968 | /* |
1969 | * unpoison_memory() can encounter thp only when the thp is being | |
1970 | * worked by memory_failure() and the page lock is not held yet. | |
1971 | * In such case, we yield to memory_failure() and make unpoison fail. | |
1972 | */ | |
e76d30e2 | 1973 | if (!PageHuge(page) && PageTransHuge(page)) { |
495367c0 | 1974 | unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n", |
a5f65109 | 1975 | pfn, &unpoison_rs); |
ead07f6a | 1976 | return 0; |
0cea3fdc WL |
1977 | } |
1978 | ||
0ed950d1 | 1979 | if (!get_hwpoison_page(p, flags)) { |
847ce401 | 1980 | if (TestClearPageHWPoison(p)) |
8e30456b | 1981 | num_poisoned_pages_dec(); |
495367c0 | 1982 | unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n", |
a5f65109 | 1983 | pfn, &unpoison_rs); |
847ce401 WF |
1984 | return 0; |
1985 | } | |
1986 | ||
7eaceacc | 1987 | lock_page(page); |
847ce401 WF |
1988 | /* |
1989 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1990 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1991 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1992 | * the free buddy page pool. | |
1993 | */ | |
c9fbdd5f | 1994 | if (TestClearPageHWPoison(page)) { |
495367c0 | 1995 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", |
a5f65109 | 1996 | pfn, &unpoison_rs); |
b37ff71c | 1997 | num_poisoned_pages_dec(); |
847ce401 WF |
1998 | freeit = 1; |
1999 | } | |
2000 | unlock_page(page); | |
2001 | ||
dd6e2402 | 2002 | put_page(page); |
3ba5eebc | 2003 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) |
dd6e2402 | 2004 | put_page(page); |
847ce401 WF |
2005 | |
2006 | return 0; | |
2007 | } | |
2008 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 2009 | |
6b9a217e | 2010 | static bool isolate_page(struct page *page, struct list_head *pagelist) |
d950b958 | 2011 | { |
6b9a217e OS |
2012 | bool isolated = false; |
2013 | bool lru = PageLRU(page); | |
d950b958 | 2014 | |
6b9a217e OS |
2015 | if (PageHuge(page)) { |
2016 | isolated = isolate_huge_page(page, pagelist); | |
2017 | } else { | |
2018 | if (lru) | |
2019 | isolated = !isolate_lru_page(page); | |
2020 | else | |
2021 | isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE); | |
2022 | ||
2023 | if (isolated) | |
2024 | list_add(&page->lru, pagelist); | |
0ebff32c | 2025 | } |
d950b958 | 2026 | |
6b9a217e OS |
2027 | if (isolated && lru) |
2028 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
2029 | page_is_file_lru(page)); | |
2030 | ||
03613808 | 2031 | /* |
6b9a217e OS |
2032 | * If we succeed to isolate the page, we grabbed another refcount on |
2033 | * the page, so we can safely drop the one we got from get_any_pages(). | |
2034 | * If we failed to isolate the page, it means that we cannot go further | |
2035 | * and we will return an error, so drop the reference we got from | |
2036 | * get_any_pages() as well. | |
03613808 | 2037 | */ |
6b9a217e OS |
2038 | put_page(page); |
2039 | return isolated; | |
d950b958 NH |
2040 | } |
2041 | ||
6b9a217e OS |
2042 | /* |
2043 | * __soft_offline_page handles hugetlb-pages and non-hugetlb pages. | |
2044 | * If the page is a non-dirty unmapped page-cache page, it simply invalidates. | |
2045 | * If the page is mapped, it migrates the contents over. | |
2046 | */ | |
2047 | static int __soft_offline_page(struct page *page) | |
af8fae7c | 2048 | { |
6b9a217e | 2049 | int ret = 0; |
af8fae7c | 2050 | unsigned long pfn = page_to_pfn(page); |
6b9a217e OS |
2051 | struct page *hpage = compound_head(page); |
2052 | char const *msg_page[] = {"page", "hugepage"}; | |
2053 | bool huge = PageHuge(page); | |
2054 | LIST_HEAD(pagelist); | |
54608759 JK |
2055 | struct migration_target_control mtc = { |
2056 | .nid = NUMA_NO_NODE, | |
2057 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
2058 | }; | |
facb6011 | 2059 | |
facb6011 | 2060 | /* |
af8fae7c NH |
2061 | * Check PageHWPoison again inside page lock because PageHWPoison |
2062 | * is set by memory_failure() outside page lock. Note that | |
2063 | * memory_failure() also double-checks PageHWPoison inside page lock, | |
2064 | * so there's no race between soft_offline_page() and memory_failure(). | |
facb6011 | 2065 | */ |
0ebff32c | 2066 | lock_page(page); |
6b9a217e OS |
2067 | if (!PageHuge(page)) |
2068 | wait_on_page_writeback(page); | |
af8fae7c NH |
2069 | if (PageHWPoison(page)) { |
2070 | unlock_page(page); | |
dd6e2402 | 2071 | put_page(page); |
af8fae7c | 2072 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
5a2ffca3 | 2073 | return 0; |
af8fae7c | 2074 | } |
6b9a217e OS |
2075 | |
2076 | if (!PageHuge(page)) | |
2077 | /* | |
2078 | * Try to invalidate first. This should work for | |
2079 | * non dirty unmapped page cache pages. | |
2080 | */ | |
2081 | ret = invalidate_inode_page(page); | |
facb6011 | 2082 | unlock_page(page); |
6b9a217e | 2083 | |
facb6011 | 2084 | /* |
facb6011 AK |
2085 | * RED-PEN would be better to keep it isolated here, but we |
2086 | * would need to fix isolation locking first. | |
2087 | */ | |
6b9a217e | 2088 | if (ret) { |
fb46e735 | 2089 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
6b9a217e | 2090 | page_handle_poison(page, false, true); |
af8fae7c | 2091 | return 0; |
facb6011 AK |
2092 | } |
2093 | ||
6b9a217e | 2094 | if (isolate_page(hpage, &pagelist)) { |
54608759 JK |
2095 | ret = migrate_pages(&pagelist, alloc_migration_target, NULL, |
2096 | (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE); | |
79f5f8fa | 2097 | if (!ret) { |
6b9a217e OS |
2098 | bool release = !huge; |
2099 | ||
2100 | if (!page_handle_poison(page, huge, release)) | |
2101 | ret = -EBUSY; | |
79f5f8fa | 2102 | } else { |
85fbe5d1 YX |
2103 | if (!list_empty(&pagelist)) |
2104 | putback_movable_pages(&pagelist); | |
59c82b70 | 2105 | |
6b9a217e OS |
2106 | pr_info("soft offline: %#lx: %s migration failed %d, type %lx (%pGp)\n", |
2107 | pfn, msg_page[huge], ret, page->flags, &page->flags); | |
facb6011 | 2108 | if (ret > 0) |
3f4b815a | 2109 | ret = -EBUSY; |
facb6011 AK |
2110 | } |
2111 | } else { | |
3f4b815a OS |
2112 | pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %lx (%pGp)\n", |
2113 | pfn, msg_page[huge], page_count(page), page->flags, &page->flags); | |
6b9a217e | 2114 | ret = -EBUSY; |
facb6011 | 2115 | } |
facb6011 AK |
2116 | return ret; |
2117 | } | |
86e05773 | 2118 | |
6b9a217e | 2119 | static int soft_offline_in_use_page(struct page *page) |
acc14dc4 | 2120 | { |
acc14dc4 NH |
2121 | struct page *hpage = compound_head(page); |
2122 | ||
694bf0b0 OS |
2123 | if (!PageHuge(page) && PageTransHuge(hpage)) |
2124 | if (try_to_split_thp_page(page, "soft offline") < 0) | |
acc14dc4 | 2125 | return -EBUSY; |
6b9a217e | 2126 | return __soft_offline_page(page); |
acc14dc4 NH |
2127 | } |
2128 | ||
d4ae9916 | 2129 | static int soft_offline_free_page(struct page *page) |
acc14dc4 | 2130 | { |
6b9a217e | 2131 | int rc = 0; |
acc14dc4 | 2132 | |
6b9a217e OS |
2133 | if (!page_handle_poison(page, true, false)) |
2134 | rc = -EBUSY; | |
06be6ff3 | 2135 | |
d4ae9916 | 2136 | return rc; |
acc14dc4 NH |
2137 | } |
2138 | ||
dad4e5b3 DW |
2139 | static void put_ref_page(struct page *page) |
2140 | { | |
2141 | if (page) | |
2142 | put_page(page); | |
2143 | } | |
2144 | ||
86e05773 WL |
2145 | /** |
2146 | * soft_offline_page - Soft offline a page. | |
feec24a6 | 2147 | * @pfn: pfn to soft-offline |
86e05773 WL |
2148 | * @flags: flags. Same as memory_failure(). |
2149 | * | |
2150 | * Returns 0 on success, otherwise negated errno. | |
2151 | * | |
2152 | * Soft offline a page, by migration or invalidation, | |
2153 | * without killing anything. This is for the case when | |
2154 | * a page is not corrupted yet (so it's still valid to access), | |
2155 | * but has had a number of corrected errors and is better taken | |
2156 | * out. | |
2157 | * | |
2158 | * The actual policy on when to do that is maintained by | |
2159 | * user space. | |
2160 | * | |
2161 | * This should never impact any application or cause data loss, | |
2162 | * however it might take some time. | |
2163 | * | |
2164 | * This is not a 100% solution for all memory, but tries to be | |
2165 | * ``good enough'' for the majority of memory. | |
2166 | */ | |
feec24a6 | 2167 | int soft_offline_page(unsigned long pfn, int flags) |
86e05773 WL |
2168 | { |
2169 | int ret; | |
b94e0282 | 2170 | bool try_again = true; |
dad4e5b3 DW |
2171 | struct page *page, *ref_page = NULL; |
2172 | ||
2173 | WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED)); | |
86e05773 | 2174 | |
feec24a6 NH |
2175 | if (!pfn_valid(pfn)) |
2176 | return -ENXIO; | |
dad4e5b3 DW |
2177 | if (flags & MF_COUNT_INCREASED) |
2178 | ref_page = pfn_to_page(pfn); | |
2179 | ||
feec24a6 NH |
2180 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ |
2181 | page = pfn_to_online_page(pfn); | |
dad4e5b3 DW |
2182 | if (!page) { |
2183 | put_ref_page(ref_page); | |
86a66810 | 2184 | return -EIO; |
dad4e5b3 | 2185 | } |
86a66810 | 2186 | |
86e05773 | 2187 | if (PageHWPoison(page)) { |
8295d535 | 2188 | pr_info("%s: %#lx page already poisoned\n", __func__, pfn); |
dad4e5b3 | 2189 | put_ref_page(ref_page); |
5a2ffca3 | 2190 | return 0; |
86e05773 | 2191 | } |
86e05773 | 2192 | |
b94e0282 | 2193 | retry: |
bfc8c901 | 2194 | get_online_mems(); |
0ed950d1 | 2195 | ret = get_hwpoison_page(page, flags); |
bfc8c901 | 2196 | put_online_mems(); |
4e41a30c | 2197 | |
8295d535 | 2198 | if (ret > 0) { |
6b9a217e | 2199 | ret = soft_offline_in_use_page(page); |
8295d535 | 2200 | } else if (ret == 0) { |
b94e0282 OS |
2201 | if (soft_offline_free_page(page) && try_again) { |
2202 | try_again = false; | |
2203 | goto retry; | |
2204 | } | |
8295d535 | 2205 | } else if (ret == -EIO) { |
6696d2a6 | 2206 | pr_info("%s: %#lx: unknown page type: %lx (%pGp)\n", |
8295d535 OS |
2207 | __func__, pfn, page->flags, &page->flags); |
2208 | } | |
4e41a30c | 2209 | |
86e05773 WL |
2210 | return ret; |
2211 | } |