]>
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> |
96c84dde | 42 | #include <linux/dax.h> |
01e00f88 | 43 | #include <linux/ksm.h> |
6a46079c | 44 | #include <linux/rmap.h> |
b9e15baf | 45 | #include <linux/export.h> |
6a46079c AK |
46 | #include <linux/pagemap.h> |
47 | #include <linux/swap.h> | |
48 | #include <linux/backing-dev.h> | |
facb6011 | 49 | #include <linux/migrate.h> |
facb6011 | 50 | #include <linux/suspend.h> |
5a0e3ad6 | 51 | #include <linux/slab.h> |
bf998156 | 52 | #include <linux/swapops.h> |
7af446a8 | 53 | #include <linux/hugetlb.h> |
20d6c96b | 54 | #include <linux/memory_hotplug.h> |
5db8a73a | 55 | #include <linux/mm_inline.h> |
6100e34b | 56 | #include <linux/memremap.h> |
ea8f5fb8 | 57 | #include <linux/kfifo.h> |
a5f65109 | 58 | #include <linux/ratelimit.h> |
d4ae9916 | 59 | #include <linux/page-isolation.h> |
a3f5d80e | 60 | #include <linux/pagewalk.h> |
a7605426 | 61 | #include <linux/shmem_fs.h> |
014bb1de | 62 | #include "swap.h" |
6a46079c | 63 | #include "internal.h" |
97f0b134 | 64 | #include "ras/ras_event.h" |
6a46079c AK |
65 | |
66 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
67 | ||
68 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
69 | ||
293c07e3 | 70 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 71 | |
67f22ba7 | 72 | static bool hw_memory_failure __read_mostly = false; |
73 | ||
510d25c9 NH |
74 | static bool __page_handle_poison(struct page *page) |
75 | { | |
f87060d3 | 76 | int ret; |
510d25c9 NH |
77 | |
78 | zone_pcp_disable(page_zone(page)); | |
79 | ret = dissolve_free_huge_page(page); | |
80 | if (!ret) | |
81 | ret = take_page_off_buddy(page); | |
82 | zone_pcp_enable(page_zone(page)); | |
83 | ||
f87060d3 | 84 | return ret > 0; |
510d25c9 NH |
85 | } |
86 | ||
6b9a217e | 87 | static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release) |
06be6ff3 | 88 | { |
6b9a217e OS |
89 | if (hugepage_or_freepage) { |
90 | /* | |
91 | * Doing this check for free pages is also fine since dissolve_free_huge_page | |
92 | * returns 0 for non-hugetlb pages as well. | |
93 | */ | |
510d25c9 | 94 | if (!__page_handle_poison(page)) |
6b9a217e OS |
95 | /* |
96 | * We could fail to take off the target page from buddy | |
f0953a1b | 97 | * for example due to racy page allocation, but that's |
6b9a217e OS |
98 | * acceptable because soft-offlined page is not broken |
99 | * and if someone really want to use it, they should | |
100 | * take it. | |
101 | */ | |
102 | return false; | |
103 | } | |
104 | ||
06be6ff3 | 105 | SetPageHWPoison(page); |
79f5f8fa OS |
106 | if (release) |
107 | put_page(page); | |
06be6ff3 OS |
108 | page_ref_inc(page); |
109 | num_poisoned_pages_inc(); | |
6b9a217e OS |
110 | |
111 | return true; | |
06be6ff3 OS |
112 | } |
113 | ||
27df5068 AK |
114 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
115 | ||
1bfe5feb | 116 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
117 | u32 hwpoison_filter_dev_major = ~0U; |
118 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
119 | u64 hwpoison_filter_flags_mask; |
120 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 121 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
122 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
123 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
124 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
125 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
126 | |
127 | static int hwpoison_filter_dev(struct page *p) | |
128 | { | |
129 | struct address_space *mapping; | |
130 | dev_t dev; | |
131 | ||
132 | if (hwpoison_filter_dev_major == ~0U && | |
133 | hwpoison_filter_dev_minor == ~0U) | |
134 | return 0; | |
135 | ||
7c116f2b WF |
136 | mapping = page_mapping(p); |
137 | if (mapping == NULL || mapping->host == NULL) | |
138 | return -EINVAL; | |
139 | ||
140 | dev = mapping->host->i_sb->s_dev; | |
141 | if (hwpoison_filter_dev_major != ~0U && | |
142 | hwpoison_filter_dev_major != MAJOR(dev)) | |
143 | return -EINVAL; | |
144 | if (hwpoison_filter_dev_minor != ~0U && | |
145 | hwpoison_filter_dev_minor != MINOR(dev)) | |
146 | return -EINVAL; | |
147 | ||
148 | return 0; | |
149 | } | |
150 | ||
478c5ffc WF |
151 | static int hwpoison_filter_flags(struct page *p) |
152 | { | |
153 | if (!hwpoison_filter_flags_mask) | |
154 | return 0; | |
155 | ||
156 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
157 | hwpoison_filter_flags_value) | |
158 | return 0; | |
159 | else | |
160 | return -EINVAL; | |
161 | } | |
162 | ||
4fd466eb AK |
163 | /* |
164 | * This allows stress tests to limit test scope to a collection of tasks | |
165 | * by putting them under some memcg. This prevents killing unrelated/important | |
166 | * processes such as /sbin/init. Note that the target task may share clean | |
167 | * pages with init (eg. libc text), which is harmless. If the target task | |
168 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
169 | * is also included in the memcg. At last, due to race conditions this filter | |
170 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
171 | * a freed page. | |
172 | */ | |
94a59fb3 | 173 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
174 | u64 hwpoison_filter_memcg; |
175 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
176 | static int hwpoison_filter_task(struct page *p) | |
177 | { | |
4fd466eb AK |
178 | if (!hwpoison_filter_memcg) |
179 | return 0; | |
180 | ||
94a59fb3 | 181 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
182 | return -EINVAL; |
183 | ||
184 | return 0; | |
185 | } | |
186 | #else | |
187 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
188 | #endif | |
189 | ||
7c116f2b WF |
190 | int hwpoison_filter(struct page *p) |
191 | { | |
1bfe5feb HL |
192 | if (!hwpoison_filter_enable) |
193 | return 0; | |
194 | ||
7c116f2b WF |
195 | if (hwpoison_filter_dev(p)) |
196 | return -EINVAL; | |
197 | ||
478c5ffc WF |
198 | if (hwpoison_filter_flags(p)) |
199 | return -EINVAL; | |
200 | ||
4fd466eb AK |
201 | if (hwpoison_filter_task(p)) |
202 | return -EINVAL; | |
203 | ||
7c116f2b WF |
204 | return 0; |
205 | } | |
27df5068 AK |
206 | #else |
207 | int hwpoison_filter(struct page *p) | |
208 | { | |
209 | return 0; | |
210 | } | |
211 | #endif | |
212 | ||
7c116f2b WF |
213 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
214 | ||
ae1139ec DW |
215 | /* |
216 | * Kill all processes that have a poisoned page mapped and then isolate | |
217 | * the page. | |
218 | * | |
219 | * General strategy: | |
220 | * Find all processes having the page mapped and kill them. | |
221 | * But we keep a page reference around so that the page is not | |
222 | * actually freed yet. | |
223 | * Then stash the page away | |
224 | * | |
225 | * There's no convenient way to get back to mapped processes | |
226 | * from the VMAs. So do a brute-force search over all | |
227 | * running processes. | |
228 | * | |
229 | * Remember that machine checks are not common (or rather | |
230 | * if they are common you have other problems), so this shouldn't | |
231 | * be a performance issue. | |
232 | * | |
233 | * Also there are some races possible while we get from the | |
234 | * error detection to actually handle it. | |
235 | */ | |
236 | ||
237 | struct to_kill { | |
238 | struct list_head nd; | |
239 | struct task_struct *tsk; | |
240 | unsigned long addr; | |
241 | short size_shift; | |
ae1139ec DW |
242 | }; |
243 | ||
6a46079c | 244 | /* |
7329bbeb TL |
245 | * Send all the processes who have the page mapped a signal. |
246 | * ``action optional'' if they are not immediately affected by the error | |
247 | * ``action required'' if error happened in current execution context | |
6a46079c | 248 | */ |
ae1139ec | 249 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 250 | { |
ae1139ec DW |
251 | struct task_struct *t = tk->tsk; |
252 | short addr_lsb = tk->size_shift; | |
872e9a20 | 253 | int ret = 0; |
6a46079c | 254 | |
03151c6e | 255 | pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", |
872e9a20 | 256 | pfn, t->comm, t->pid); |
7329bbeb | 257 | |
49775047 ML |
258 | if ((flags & MF_ACTION_REQUIRED) && (t == current)) |
259 | ret = force_sig_mceerr(BUS_MCEERR_AR, | |
260 | (void __user *)tk->addr, addr_lsb); | |
261 | else | |
7329bbeb | 262 | /* |
49775047 ML |
263 | * Signal other processes sharing the page if they have |
264 | * PF_MCE_EARLY set. | |
7329bbeb TL |
265 | * Don't use force here, it's convenient if the signal |
266 | * can be temporarily blocked. | |
267 | * This could cause a loop when the user sets SIGBUS | |
268 | * to SIG_IGN, but hopefully no one will do that? | |
269 | */ | |
ae1139ec | 270 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
c0f45555 | 271 | addr_lsb, t); /* synchronous? */ |
6a46079c | 272 | if (ret < 0) |
495367c0 | 273 | pr_info("Memory failure: Error sending signal to %s:%d: %d\n", |
1170532b | 274 | t->comm, t->pid, ret); |
6a46079c AK |
275 | return ret; |
276 | } | |
277 | ||
588f9ce6 | 278 | /* |
47e431f4 | 279 | * Unknown page type encountered. Try to check whether it can turn PageLRU by |
d0505e9f | 280 | * lru_add_drain_all. |
588f9ce6 | 281 | */ |
d0505e9f | 282 | void shake_page(struct page *p) |
588f9ce6 | 283 | { |
8bcb74de NH |
284 | if (PageHuge(p)) |
285 | return; | |
286 | ||
588f9ce6 AK |
287 | if (!PageSlab(p)) { |
288 | lru_add_drain_all(); | |
588f9ce6 AK |
289 | if (PageLRU(p) || is_free_buddy_page(p)) |
290 | return; | |
291 | } | |
facb6011 | 292 | |
588f9ce6 | 293 | /* |
d0505e9f YS |
294 | * TODO: Could shrink slab caches here if a lightweight range-based |
295 | * shrinker will be available. | |
588f9ce6 AK |
296 | */ |
297 | } | |
298 | EXPORT_SYMBOL_GPL(shake_page); | |
299 | ||
6100e34b DW |
300 | static unsigned long dev_pagemap_mapping_shift(struct page *page, |
301 | struct vm_area_struct *vma) | |
302 | { | |
303 | unsigned long address = vma_address(page, vma); | |
5c91c0e7 | 304 | unsigned long ret = 0; |
6100e34b DW |
305 | pgd_t *pgd; |
306 | p4d_t *p4d; | |
307 | pud_t *pud; | |
308 | pmd_t *pmd; | |
309 | pte_t *pte; | |
310 | ||
a994402b | 311 | VM_BUG_ON_VMA(address == -EFAULT, vma); |
6100e34b DW |
312 | pgd = pgd_offset(vma->vm_mm, address); |
313 | if (!pgd_present(*pgd)) | |
314 | return 0; | |
315 | p4d = p4d_offset(pgd, address); | |
316 | if (!p4d_present(*p4d)) | |
317 | return 0; | |
318 | pud = pud_offset(p4d, address); | |
319 | if (!pud_present(*pud)) | |
320 | return 0; | |
321 | if (pud_devmap(*pud)) | |
322 | return PUD_SHIFT; | |
323 | pmd = pmd_offset(pud, address); | |
324 | if (!pmd_present(*pmd)) | |
325 | return 0; | |
326 | if (pmd_devmap(*pmd)) | |
327 | return PMD_SHIFT; | |
328 | pte = pte_offset_map(pmd, address); | |
5c91c0e7 QZ |
329 | if (pte_present(*pte) && pte_devmap(*pte)) |
330 | ret = PAGE_SHIFT; | |
331 | pte_unmap(pte); | |
332 | return ret; | |
6100e34b | 333 | } |
6a46079c AK |
334 | |
335 | /* | |
336 | * Failure handling: if we can't find or can't kill a process there's | |
337 | * not much we can do. We just print a message and ignore otherwise. | |
338 | */ | |
339 | ||
340 | /* | |
341 | * Schedule a process for later kill. | |
342 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
6a46079c AK |
343 | */ |
344 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
345 | struct vm_area_struct *vma, | |
996ff7a0 | 346 | struct list_head *to_kill) |
6a46079c AK |
347 | { |
348 | struct to_kill *tk; | |
349 | ||
996ff7a0 JC |
350 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
351 | if (!tk) { | |
352 | pr_err("Memory failure: Out of memory while machine check handling\n"); | |
353 | return; | |
6a46079c | 354 | } |
996ff7a0 | 355 | |
6a46079c | 356 | tk->addr = page_address_in_vma(p, vma); |
6100e34b DW |
357 | if (is_zone_device_page(p)) |
358 | tk->size_shift = dev_pagemap_mapping_shift(p, vma); | |
359 | else | |
75068518 | 360 | tk->size_shift = page_shift(compound_head(p)); |
6a46079c AK |
361 | |
362 | /* | |
3d7fed4a JC |
363 | * Send SIGKILL if "tk->addr == -EFAULT". Also, as |
364 | * "tk->size_shift" is always non-zero for !is_zone_device_page(), | |
365 | * so "tk->size_shift == 0" effectively checks no mapping on | |
366 | * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times | |
367 | * to a process' address space, it's possible not all N VMAs | |
368 | * contain mappings for the page, but at least one VMA does. | |
369 | * Only deliver SIGBUS with payload derived from the VMA that | |
370 | * has a mapping for the page. | |
6a46079c | 371 | */ |
3d7fed4a | 372 | if (tk->addr == -EFAULT) { |
495367c0 | 373 | pr_info("Memory failure: Unable to find user space address %lx in %s\n", |
6a46079c | 374 | page_to_pfn(p), tsk->comm); |
3d7fed4a JC |
375 | } else if (tk->size_shift == 0) { |
376 | kfree(tk); | |
377 | return; | |
6a46079c | 378 | } |
996ff7a0 | 379 | |
6a46079c AK |
380 | get_task_struct(tsk); |
381 | tk->tsk = tsk; | |
382 | list_add_tail(&tk->nd, to_kill); | |
383 | } | |
384 | ||
385 | /* | |
386 | * Kill the processes that have been collected earlier. | |
387 | * | |
a21c184f ML |
388 | * Only do anything when FORCEKILL is set, otherwise just free the |
389 | * list (this is used for clean pages which do not need killing) | |
6a46079c AK |
390 | * Also when FAIL is set do a force kill because something went |
391 | * wrong earlier. | |
392 | */ | |
ae1139ec DW |
393 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
394 | unsigned long pfn, int flags) | |
6a46079c AK |
395 | { |
396 | struct to_kill *tk, *next; | |
397 | ||
398 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 399 | if (forcekill) { |
6a46079c | 400 | /* |
af901ca1 | 401 | * In case something went wrong with munmapping |
6a46079c AK |
402 | * make sure the process doesn't catch the |
403 | * signal and then access the memory. Just kill it. | |
6a46079c | 404 | */ |
3d7fed4a | 405 | if (fail || tk->addr == -EFAULT) { |
495367c0 | 406 | pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 407 | pfn, tk->tsk->comm, tk->tsk->pid); |
6376360e NH |
408 | do_send_sig_info(SIGKILL, SEND_SIG_PRIV, |
409 | tk->tsk, PIDTYPE_PID); | |
6a46079c AK |
410 | } |
411 | ||
412 | /* | |
413 | * In theory the process could have mapped | |
414 | * something else on the address in-between. We could | |
415 | * check for that, but we need to tell the | |
416 | * process anyways. | |
417 | */ | |
ae1139ec | 418 | else if (kill_proc(tk, pfn, flags) < 0) |
495367c0 | 419 | pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 420 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c AK |
421 | } |
422 | put_task_struct(tk->tsk); | |
423 | kfree(tk); | |
424 | } | |
425 | } | |
426 | ||
3ba08129 NH |
427 | /* |
428 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
429 | * on behalf of the thread group. Return task_struct of the (first found) | |
430 | * dedicated thread if found, and return NULL otherwise. | |
431 | * | |
432 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
433 | * have to call rcu_read_lock/unlock() in this function. | |
434 | */ | |
435 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 436 | { |
3ba08129 NH |
437 | struct task_struct *t; |
438 | ||
4e018b45 NH |
439 | for_each_thread(tsk, t) { |
440 | if (t->flags & PF_MCE_PROCESS) { | |
441 | if (t->flags & PF_MCE_EARLY) | |
442 | return t; | |
443 | } else { | |
444 | if (sysctl_memory_failure_early_kill) | |
445 | return t; | |
446 | } | |
447 | } | |
3ba08129 NH |
448 | return NULL; |
449 | } | |
450 | ||
451 | /* | |
452 | * Determine whether a given process is "early kill" process which expects | |
453 | * to be signaled when some page under the process is hwpoisoned. | |
454 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
30c9cf49 | 455 | * specified) if the process is "early kill" and otherwise returns NULL. |
03151c6e | 456 | * |
30c9cf49 AY |
457 | * Note that the above is true for Action Optional case. For Action Required |
458 | * case, it's only meaningful to the current thread which need to be signaled | |
459 | * with SIGBUS, this error is Action Optional for other non current | |
460 | * processes sharing the same error page,if the process is "early kill", the | |
461 | * task_struct of the dedicated thread will also be returned. | |
3ba08129 NH |
462 | */ |
463 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
464 | int force_early) | |
465 | { | |
6a46079c | 466 | if (!tsk->mm) |
3ba08129 | 467 | return NULL; |
30c9cf49 AY |
468 | /* |
469 | * Comparing ->mm here because current task might represent | |
470 | * a subthread, while tsk always points to the main thread. | |
471 | */ | |
472 | if (force_early && tsk->mm == current->mm) | |
473 | return current; | |
474 | ||
4e018b45 | 475 | return find_early_kill_thread(tsk); |
6a46079c AK |
476 | } |
477 | ||
478 | /* | |
479 | * Collect processes when the error hit an anonymous page. | |
480 | */ | |
481 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 482 | int force_early) |
6a46079c | 483 | { |
9595d769 | 484 | struct folio *folio = page_folio(page); |
6a46079c AK |
485 | struct vm_area_struct *vma; |
486 | struct task_struct *tsk; | |
487 | struct anon_vma *av; | |
bf181b9f | 488 | pgoff_t pgoff; |
6a46079c | 489 | |
6d4675e6 | 490 | av = folio_lock_anon_vma_read(folio, NULL); |
6a46079c | 491 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
492 | return; |
493 | ||
a0f7a756 | 494 | pgoff = page_to_pgoff(page); |
9b679320 | 495 | read_lock(&tasklist_lock); |
6a46079c | 496 | for_each_process (tsk) { |
5beb4930 | 497 | struct anon_vma_chain *vmac; |
3ba08129 | 498 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 499 | |
3ba08129 | 500 | if (!t) |
6a46079c | 501 | continue; |
bf181b9f ML |
502 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
503 | pgoff, pgoff) { | |
5beb4930 | 504 | vma = vmac->vma; |
6a46079c AK |
505 | if (!page_mapped_in_vma(page, vma)) |
506 | continue; | |
3ba08129 | 507 | if (vma->vm_mm == t->mm) |
996ff7a0 | 508 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
509 | } |
510 | } | |
6a46079c | 511 | read_unlock(&tasklist_lock); |
4fc3f1d6 | 512 | page_unlock_anon_vma_read(av); |
6a46079c AK |
513 | } |
514 | ||
515 | /* | |
516 | * Collect processes when the error hit a file mapped page. | |
517 | */ | |
518 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 519 | int force_early) |
6a46079c AK |
520 | { |
521 | struct vm_area_struct *vma; | |
522 | struct task_struct *tsk; | |
6a46079c | 523 | struct address_space *mapping = page->mapping; |
c43bc03d | 524 | pgoff_t pgoff; |
6a46079c | 525 | |
d28eb9c8 | 526 | i_mmap_lock_read(mapping); |
9b679320 | 527 | read_lock(&tasklist_lock); |
c43bc03d | 528 | pgoff = page_to_pgoff(page); |
6a46079c | 529 | for_each_process(tsk) { |
3ba08129 | 530 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 531 | |
3ba08129 | 532 | if (!t) |
6a46079c | 533 | continue; |
6b2dbba8 | 534 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
535 | pgoff) { |
536 | /* | |
537 | * Send early kill signal to tasks where a vma covers | |
538 | * the page but the corrupted page is not necessarily | |
539 | * mapped it in its pte. | |
540 | * Assume applications who requested early kill want | |
541 | * to be informed of all such data corruptions. | |
542 | */ | |
3ba08129 | 543 | if (vma->vm_mm == t->mm) |
996ff7a0 | 544 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
545 | } |
546 | } | |
6a46079c | 547 | read_unlock(&tasklist_lock); |
d28eb9c8 | 548 | i_mmap_unlock_read(mapping); |
6a46079c AK |
549 | } |
550 | ||
551 | /* | |
552 | * Collect the processes who have the corrupted page mapped to kill. | |
6a46079c | 553 | */ |
74614de1 TL |
554 | static void collect_procs(struct page *page, struct list_head *tokill, |
555 | int force_early) | |
6a46079c | 556 | { |
6a46079c AK |
557 | if (!page->mapping) |
558 | return; | |
559 | ||
6a46079c | 560 | if (PageAnon(page)) |
996ff7a0 | 561 | collect_procs_anon(page, tokill, force_early); |
6a46079c | 562 | else |
996ff7a0 | 563 | collect_procs_file(page, tokill, force_early); |
6a46079c AK |
564 | } |
565 | ||
a3f5d80e NH |
566 | struct hwp_walk { |
567 | struct to_kill tk; | |
568 | unsigned long pfn; | |
569 | int flags; | |
570 | }; | |
571 | ||
572 | static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift) | |
573 | { | |
574 | tk->addr = addr; | |
575 | tk->size_shift = shift; | |
576 | } | |
577 | ||
578 | static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift, | |
579 | unsigned long poisoned_pfn, struct to_kill *tk) | |
580 | { | |
581 | unsigned long pfn = 0; | |
582 | ||
583 | if (pte_present(pte)) { | |
584 | pfn = pte_pfn(pte); | |
585 | } else { | |
586 | swp_entry_t swp = pte_to_swp_entry(pte); | |
587 | ||
588 | if (is_hwpoison_entry(swp)) | |
589 | pfn = hwpoison_entry_to_pfn(swp); | |
590 | } | |
591 | ||
592 | if (!pfn || pfn != poisoned_pfn) | |
593 | return 0; | |
594 | ||
595 | set_to_kill(tk, addr, shift); | |
596 | return 1; | |
597 | } | |
598 | ||
599 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
600 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
601 | struct hwp_walk *hwp) | |
602 | { | |
603 | pmd_t pmd = *pmdp; | |
604 | unsigned long pfn; | |
605 | unsigned long hwpoison_vaddr; | |
606 | ||
607 | if (!pmd_present(pmd)) | |
608 | return 0; | |
609 | pfn = pmd_pfn(pmd); | |
610 | if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) { | |
611 | hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT); | |
612 | set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT); | |
613 | return 1; | |
614 | } | |
615 | return 0; | |
616 | } | |
617 | #else | |
618 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
619 | struct hwp_walk *hwp) | |
620 | { | |
621 | return 0; | |
622 | } | |
623 | #endif | |
624 | ||
625 | static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr, | |
626 | unsigned long end, struct mm_walk *walk) | |
627 | { | |
f142e707 | 628 | struct hwp_walk *hwp = walk->private; |
a3f5d80e | 629 | int ret = 0; |
ea3732f7 | 630 | pte_t *ptep, *mapped_pte; |
a3f5d80e NH |
631 | spinlock_t *ptl; |
632 | ||
633 | ptl = pmd_trans_huge_lock(pmdp, walk->vma); | |
634 | if (ptl) { | |
635 | ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp); | |
636 | spin_unlock(ptl); | |
637 | goto out; | |
638 | } | |
639 | ||
640 | if (pmd_trans_unstable(pmdp)) | |
641 | goto out; | |
642 | ||
ea3732f7 ML |
643 | mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, |
644 | addr, &ptl); | |
a3f5d80e NH |
645 | for (; addr != end; ptep++, addr += PAGE_SIZE) { |
646 | ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT, | |
647 | hwp->pfn, &hwp->tk); | |
648 | if (ret == 1) | |
649 | break; | |
650 | } | |
ea3732f7 | 651 | pte_unmap_unlock(mapped_pte, ptl); |
a3f5d80e NH |
652 | out: |
653 | cond_resched(); | |
654 | return ret; | |
655 | } | |
656 | ||
657 | #ifdef CONFIG_HUGETLB_PAGE | |
658 | static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask, | |
659 | unsigned long addr, unsigned long end, | |
660 | struct mm_walk *walk) | |
661 | { | |
f142e707 | 662 | struct hwp_walk *hwp = walk->private; |
a3f5d80e NH |
663 | pte_t pte = huge_ptep_get(ptep); |
664 | struct hstate *h = hstate_vma(walk->vma); | |
665 | ||
666 | return check_hwpoisoned_entry(pte, addr, huge_page_shift(h), | |
667 | hwp->pfn, &hwp->tk); | |
668 | } | |
669 | #else | |
670 | #define hwpoison_hugetlb_range NULL | |
671 | #endif | |
672 | ||
ba9eb3ce | 673 | static const struct mm_walk_ops hwp_walk_ops = { |
a3f5d80e NH |
674 | .pmd_entry = hwpoison_pte_range, |
675 | .hugetlb_entry = hwpoison_hugetlb_range, | |
676 | }; | |
677 | ||
678 | /* | |
679 | * Sends SIGBUS to the current process with error info. | |
680 | * | |
681 | * This function is intended to handle "Action Required" MCEs on already | |
682 | * hardware poisoned pages. They could happen, for example, when | |
683 | * memory_failure() failed to unmap the error page at the first call, or | |
684 | * when multiple local machine checks happened on different CPUs. | |
685 | * | |
686 | * MCE handler currently has no easy access to the error virtual address, | |
687 | * so this function walks page table to find it. The returned virtual address | |
688 | * is proper in most cases, but it could be wrong when the application | |
689 | * process has multiple entries mapping the error page. | |
690 | */ | |
691 | static int kill_accessing_process(struct task_struct *p, unsigned long pfn, | |
692 | int flags) | |
693 | { | |
694 | int ret; | |
695 | struct hwp_walk priv = { | |
696 | .pfn = pfn, | |
697 | }; | |
698 | priv.tk.tsk = p; | |
699 | ||
700 | mmap_read_lock(p->mm); | |
701 | ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops, | |
702 | (void *)&priv); | |
703 | if (ret == 1 && priv.tk.addr) | |
704 | kill_proc(&priv.tk, pfn, flags); | |
046545a6 NH |
705 | else |
706 | ret = 0; | |
a3f5d80e | 707 | mmap_read_unlock(p->mm); |
046545a6 | 708 | return ret > 0 ? -EHWPOISON : -EFAULT; |
a3f5d80e NH |
709 | } |
710 | ||
6a46079c | 711 | static const char *action_name[] = { |
cc637b17 XX |
712 | [MF_IGNORED] = "Ignored", |
713 | [MF_FAILED] = "Failed", | |
714 | [MF_DELAYED] = "Delayed", | |
715 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
716 | }; |
717 | ||
718 | static const char * const action_page_types[] = { | |
cc637b17 XX |
719 | [MF_MSG_KERNEL] = "reserved kernel page", |
720 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
721 | [MF_MSG_SLAB] = "kernel slab page", | |
722 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
cc637b17 XX |
723 | [MF_MSG_HUGE] = "huge page", |
724 | [MF_MSG_FREE_HUGE] = "free huge page", | |
31286a84 | 725 | [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page", |
cc637b17 XX |
726 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", |
727 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
728 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
729 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
730 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
731 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
732 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
733 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
734 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
735 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
736 | [MF_MSG_BUDDY] = "free buddy page", | |
6100e34b | 737 | [MF_MSG_DAX] = "dax page", |
5d1fd5dc | 738 | [MF_MSG_UNSPLIT_THP] = "unsplit thp", |
cc637b17 | 739 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
740 | }; |
741 | ||
dc2a1cbf WF |
742 | /* |
743 | * XXX: It is possible that a page is isolated from LRU cache, | |
744 | * and then kept in swap cache or failed to remove from page cache. | |
745 | * The page count will stop it from being freed by unpoison. | |
746 | * Stress tests should be aware of this memory leak problem. | |
747 | */ | |
748 | static int delete_from_lru_cache(struct page *p) | |
749 | { | |
750 | if (!isolate_lru_page(p)) { | |
751 | /* | |
752 | * Clear sensible page flags, so that the buddy system won't | |
753 | * complain when the page is unpoison-and-freed. | |
754 | */ | |
755 | ClearPageActive(p); | |
756 | ClearPageUnevictable(p); | |
18365225 MH |
757 | |
758 | /* | |
759 | * Poisoned page might never drop its ref count to 0 so we have | |
760 | * to uncharge it manually from its memcg. | |
761 | */ | |
bbc6b703 | 762 | mem_cgroup_uncharge(page_folio(p)); |
18365225 | 763 | |
dc2a1cbf WF |
764 | /* |
765 | * drop the page count elevated by isolate_lru_page() | |
766 | */ | |
09cbfeaf | 767 | put_page(p); |
dc2a1cbf WF |
768 | return 0; |
769 | } | |
770 | return -EIO; | |
771 | } | |
772 | ||
78bb9203 NH |
773 | static int truncate_error_page(struct page *p, unsigned long pfn, |
774 | struct address_space *mapping) | |
775 | { | |
776 | int ret = MF_FAILED; | |
777 | ||
778 | if (mapping->a_ops->error_remove_page) { | |
779 | int err = mapping->a_ops->error_remove_page(mapping, p); | |
780 | ||
781 | if (err != 0) { | |
782 | pr_info("Memory failure: %#lx: Failed to punch page: %d\n", | |
783 | pfn, err); | |
784 | } else if (page_has_private(p) && | |
785 | !try_to_release_page(p, GFP_NOIO)) { | |
786 | pr_info("Memory failure: %#lx: failed to release buffers\n", | |
787 | pfn); | |
788 | } else { | |
789 | ret = MF_RECOVERED; | |
790 | } | |
791 | } else { | |
792 | /* | |
793 | * If the file system doesn't support it just invalidate | |
794 | * This fails on dirty or anything with private pages | |
795 | */ | |
796 | if (invalidate_inode_page(p)) | |
797 | ret = MF_RECOVERED; | |
798 | else | |
799 | pr_info("Memory failure: %#lx: Failed to invalidate\n", | |
800 | pfn); | |
801 | } | |
802 | ||
803 | return ret; | |
804 | } | |
805 | ||
dd0f230a YS |
806 | struct page_state { |
807 | unsigned long mask; | |
808 | unsigned long res; | |
809 | enum mf_action_page_type type; | |
810 | ||
811 | /* Callback ->action() has to unlock the relevant page inside it. */ | |
812 | int (*action)(struct page_state *ps, struct page *p); | |
813 | }; | |
814 | ||
815 | /* | |
816 | * Return true if page is still referenced by others, otherwise return | |
817 | * false. | |
818 | * | |
819 | * The extra_pins is true when one extra refcount is expected. | |
820 | */ | |
821 | static bool has_extra_refcount(struct page_state *ps, struct page *p, | |
822 | bool extra_pins) | |
823 | { | |
824 | int count = page_count(p) - 1; | |
825 | ||
826 | if (extra_pins) | |
827 | count -= 1; | |
828 | ||
829 | if (count > 0) { | |
830 | pr_err("Memory failure: %#lx: %s still referenced by %d users\n", | |
831 | page_to_pfn(p), action_page_types[ps->type], count); | |
832 | return true; | |
833 | } | |
834 | ||
835 | return false; | |
836 | } | |
837 | ||
6a46079c AK |
838 | /* |
839 | * Error hit kernel page. | |
840 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
841 | * could be more sophisticated. | |
842 | */ | |
dd0f230a | 843 | static int me_kernel(struct page_state *ps, struct page *p) |
6a46079c | 844 | { |
ea6d0630 | 845 | unlock_page(p); |
cc637b17 | 846 | return MF_IGNORED; |
6a46079c AK |
847 | } |
848 | ||
849 | /* | |
850 | * Page in unknown state. Do nothing. | |
851 | */ | |
dd0f230a | 852 | static int me_unknown(struct page_state *ps, struct page *p) |
6a46079c | 853 | { |
dd0f230a | 854 | pr_err("Memory failure: %#lx: Unknown page state\n", page_to_pfn(p)); |
ea6d0630 | 855 | unlock_page(p); |
cc637b17 | 856 | return MF_FAILED; |
6a46079c AK |
857 | } |
858 | ||
6a46079c AK |
859 | /* |
860 | * Clean (or cleaned) page cache page. | |
861 | */ | |
dd0f230a | 862 | static int me_pagecache_clean(struct page_state *ps, struct page *p) |
6a46079c | 863 | { |
ea6d0630 | 864 | int ret; |
6a46079c | 865 | struct address_space *mapping; |
a7605426 | 866 | bool extra_pins; |
6a46079c | 867 | |
dc2a1cbf WF |
868 | delete_from_lru_cache(p); |
869 | ||
6a46079c AK |
870 | /* |
871 | * For anonymous pages we're done the only reference left | |
872 | * should be the one m_f() holds. | |
873 | */ | |
ea6d0630 NH |
874 | if (PageAnon(p)) { |
875 | ret = MF_RECOVERED; | |
876 | goto out; | |
877 | } | |
6a46079c AK |
878 | |
879 | /* | |
880 | * Now truncate the page in the page cache. This is really | |
881 | * more like a "temporary hole punch" | |
882 | * Don't do this for block devices when someone else | |
883 | * has a reference, because it could be file system metadata | |
884 | * and that's not safe to truncate. | |
885 | */ | |
886 | mapping = page_mapping(p); | |
887 | if (!mapping) { | |
888 | /* | |
889 | * Page has been teared down in the meanwhile | |
890 | */ | |
ea6d0630 NH |
891 | ret = MF_FAILED; |
892 | goto out; | |
6a46079c AK |
893 | } |
894 | ||
a7605426 YS |
895 | /* |
896 | * The shmem page is kept in page cache instead of truncating | |
897 | * so is expected to have an extra refcount after error-handling. | |
898 | */ | |
899 | extra_pins = shmem_mapping(mapping); | |
900 | ||
6a46079c AK |
901 | /* |
902 | * Truncation is a bit tricky. Enable it per file system for now. | |
903 | * | |
9608703e | 904 | * Open: to take i_rwsem or not for this? Right now we don't. |
6a46079c | 905 | */ |
dd0f230a | 906 | ret = truncate_error_page(p, page_to_pfn(p), mapping); |
a7605426 YS |
907 | if (has_extra_refcount(ps, p, extra_pins)) |
908 | ret = MF_FAILED; | |
909 | ||
ea6d0630 NH |
910 | out: |
911 | unlock_page(p); | |
dd0f230a | 912 | |
ea6d0630 | 913 | return ret; |
6a46079c AK |
914 | } |
915 | ||
916 | /* | |
549543df | 917 | * Dirty pagecache page |
6a46079c AK |
918 | * Issues: when the error hit a hole page the error is not properly |
919 | * propagated. | |
920 | */ | |
dd0f230a | 921 | static int me_pagecache_dirty(struct page_state *ps, struct page *p) |
6a46079c AK |
922 | { |
923 | struct address_space *mapping = page_mapping(p); | |
924 | ||
925 | SetPageError(p); | |
926 | /* TBD: print more information about the file. */ | |
927 | if (mapping) { | |
928 | /* | |
929 | * IO error will be reported by write(), fsync(), etc. | |
930 | * who check the mapping. | |
931 | * This way the application knows that something went | |
932 | * wrong with its dirty file data. | |
933 | * | |
934 | * There's one open issue: | |
935 | * | |
936 | * The EIO will be only reported on the next IO | |
937 | * operation and then cleared through the IO map. | |
938 | * Normally Linux has two mechanisms to pass IO error | |
939 | * first through the AS_EIO flag in the address space | |
940 | * and then through the PageError flag in the page. | |
941 | * Since we drop pages on memory failure handling the | |
942 | * only mechanism open to use is through AS_AIO. | |
943 | * | |
944 | * This has the disadvantage that it gets cleared on | |
945 | * the first operation that returns an error, while | |
946 | * the PageError bit is more sticky and only cleared | |
947 | * when the page is reread or dropped. If an | |
948 | * application assumes it will always get error on | |
949 | * fsync, but does other operations on the fd before | |
25985edc | 950 | * and the page is dropped between then the error |
6a46079c AK |
951 | * will not be properly reported. |
952 | * | |
953 | * This can already happen even without hwpoisoned | |
954 | * pages: first on metadata IO errors (which only | |
955 | * report through AS_EIO) or when the page is dropped | |
956 | * at the wrong time. | |
957 | * | |
958 | * So right now we assume that the application DTRT on | |
959 | * the first EIO, but we're not worse than other parts | |
960 | * of the kernel. | |
961 | */ | |
af21bfaf | 962 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
963 | } |
964 | ||
dd0f230a | 965 | return me_pagecache_clean(ps, p); |
6a46079c AK |
966 | } |
967 | ||
968 | /* | |
969 | * Clean and dirty swap cache. | |
970 | * | |
971 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
972 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
973 | * referenced concurrently by 2 types of PTEs: | |
974 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
975 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
976 | * and then | |
977 | * - clear dirty bit to prevent IO | |
978 | * - remove from LRU | |
979 | * - but keep in the swap cache, so that when we return to it on | |
980 | * a later page fault, we know the application is accessing | |
981 | * corrupted data and shall be killed (we installed simple | |
982 | * interception code in do_swap_page to catch it). | |
983 | * | |
984 | * Clean swap cache pages can be directly isolated. A later page fault will | |
985 | * bring in the known good data from disk. | |
986 | */ | |
dd0f230a | 987 | static int me_swapcache_dirty(struct page_state *ps, struct page *p) |
6a46079c | 988 | { |
ea6d0630 | 989 | int ret; |
dd0f230a | 990 | bool extra_pins = false; |
ea6d0630 | 991 | |
6a46079c AK |
992 | ClearPageDirty(p); |
993 | /* Trigger EIO in shmem: */ | |
994 | ClearPageUptodate(p); | |
995 | ||
ea6d0630 NH |
996 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED; |
997 | unlock_page(p); | |
dd0f230a YS |
998 | |
999 | if (ret == MF_DELAYED) | |
1000 | extra_pins = true; | |
1001 | ||
1002 | if (has_extra_refcount(ps, p, extra_pins)) | |
1003 | ret = MF_FAILED; | |
1004 | ||
ea6d0630 | 1005 | return ret; |
6a46079c AK |
1006 | } |
1007 | ||
dd0f230a | 1008 | static int me_swapcache_clean(struct page_state *ps, struct page *p) |
6a46079c | 1009 | { |
ea6d0630 NH |
1010 | int ret; |
1011 | ||
6a46079c | 1012 | delete_from_swap_cache(p); |
e43c3afb | 1013 | |
ea6d0630 NH |
1014 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED; |
1015 | unlock_page(p); | |
dd0f230a YS |
1016 | |
1017 | if (has_extra_refcount(ps, p, false)) | |
1018 | ret = MF_FAILED; | |
1019 | ||
ea6d0630 | 1020 | return ret; |
6a46079c AK |
1021 | } |
1022 | ||
1023 | /* | |
1024 | * Huge pages. Needs work. | |
1025 | * Issues: | |
93f70f90 NH |
1026 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
1027 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c | 1028 | */ |
dd0f230a | 1029 | static int me_huge_page(struct page_state *ps, struct page *p) |
6a46079c | 1030 | { |
a8b2c2ce | 1031 | int res; |
93f70f90 | 1032 | struct page *hpage = compound_head(p); |
78bb9203 | 1033 | struct address_space *mapping; |
2491ffee NH |
1034 | |
1035 | if (!PageHuge(hpage)) | |
1036 | return MF_DELAYED; | |
1037 | ||
78bb9203 NH |
1038 | mapping = page_mapping(hpage); |
1039 | if (mapping) { | |
dd0f230a | 1040 | res = truncate_error_page(hpage, page_to_pfn(p), mapping); |
ea6d0630 | 1041 | unlock_page(hpage); |
78bb9203 | 1042 | } else { |
a8b2c2ce | 1043 | res = MF_FAILED; |
78bb9203 NH |
1044 | unlock_page(hpage); |
1045 | /* | |
ef526b17 ML |
1046 | * migration entry prevents later access on error hugepage, |
1047 | * so we can free and dissolve it into buddy to save healthy | |
1048 | * subpages. | |
78bb9203 | 1049 | */ |
ef526b17 | 1050 | put_page(hpage); |
510d25c9 | 1051 | if (__page_handle_poison(p)) { |
a8b2c2ce OS |
1052 | page_ref_inc(p); |
1053 | res = MF_RECOVERED; | |
1054 | } | |
93f70f90 | 1055 | } |
78bb9203 | 1056 | |
dd0f230a YS |
1057 | if (has_extra_refcount(ps, p, false)) |
1058 | res = MF_FAILED; | |
1059 | ||
78bb9203 | 1060 | return res; |
6a46079c AK |
1061 | } |
1062 | ||
1063 | /* | |
1064 | * Various page states we can handle. | |
1065 | * | |
1066 | * A page state is defined by its current page->flags bits. | |
1067 | * The table matches them in order and calls the right handler. | |
1068 | * | |
1069 | * This is quite tricky because we can access page at any time | |
25985edc | 1070 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
1071 | * |
1072 | * This is not complete. More states could be added. | |
1073 | * For any missing state don't attempt recovery. | |
1074 | */ | |
1075 | ||
1076 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 1077 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
1078 | #define unevict (1UL << PG_unevictable) |
1079 | #define mlock (1UL << PG_mlocked) | |
6a46079c | 1080 | #define lru (1UL << PG_lru) |
6a46079c | 1081 | #define head (1UL << PG_head) |
6a46079c | 1082 | #define slab (1UL << PG_slab) |
6a46079c AK |
1083 | #define reserved (1UL << PG_reserved) |
1084 | ||
dd0f230a | 1085 | static struct page_state error_states[] = { |
cc637b17 | 1086 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
1087 | /* |
1088 | * free pages are specially detected outside this table: | |
1089 | * PG_buddy pages only make a small fraction of all free pages. | |
1090 | */ | |
6a46079c AK |
1091 | |
1092 | /* | |
1093 | * Could in theory check if slab page is free or if we can drop | |
1094 | * currently unused objects without touching them. But just | |
1095 | * treat it as standard kernel for now. | |
1096 | */ | |
cc637b17 | 1097 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 1098 | |
cc637b17 | 1099 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 1100 | |
cc637b17 XX |
1101 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
1102 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 1103 | |
cc637b17 XX |
1104 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
1105 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 1106 | |
cc637b17 XX |
1107 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
1108 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 1109 | |
cc637b17 XX |
1110 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
1111 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
1112 | |
1113 | /* | |
1114 | * Catchall entry: must be at end. | |
1115 | */ | |
cc637b17 | 1116 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
1117 | }; |
1118 | ||
2326c467 AK |
1119 | #undef dirty |
1120 | #undef sc | |
1121 | #undef unevict | |
1122 | #undef mlock | |
2326c467 | 1123 | #undef lru |
2326c467 | 1124 | #undef head |
2326c467 AK |
1125 | #undef slab |
1126 | #undef reserved | |
1127 | ||
ff604cf6 NH |
1128 | /* |
1129 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
1130 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
1131 | */ | |
cc3e2af4 XX |
1132 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
1133 | enum mf_result result) | |
6a46079c | 1134 | { |
97f0b134 XX |
1135 | trace_memory_failure_event(pfn, type, result); |
1136 | ||
e240ac52 | 1137 | num_poisoned_pages_inc(); |
495367c0 | 1138 | pr_err("Memory failure: %#lx: recovery action for %s: %s\n", |
64d37a2b | 1139 | pfn, action_page_types[type], action_name[result]); |
6a46079c AK |
1140 | } |
1141 | ||
1142 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 1143 | unsigned long pfn) |
6a46079c AK |
1144 | { |
1145 | int result; | |
1146 | ||
ea6d0630 | 1147 | /* page p should be unlocked after returning from ps->action(). */ |
dd0f230a | 1148 | result = ps->action(ps, p); |
7456b040 | 1149 | |
64d37a2b | 1150 | action_result(pfn, ps->type, result); |
6a46079c AK |
1151 | |
1152 | /* Could do more checks here if page looks ok */ | |
1153 | /* | |
1154 | * Could adjust zone counters here to correct for the missing page. | |
1155 | */ | |
1156 | ||
cc637b17 | 1157 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
1158 | } |
1159 | ||
bf181c58 NH |
1160 | static inline bool PageHWPoisonTakenOff(struct page *page) |
1161 | { | |
1162 | return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON; | |
1163 | } | |
1164 | ||
1165 | void SetPageHWPoisonTakenOff(struct page *page) | |
1166 | { | |
1167 | set_page_private(page, MAGIC_HWPOISON); | |
1168 | } | |
1169 | ||
1170 | void ClearPageHWPoisonTakenOff(struct page *page) | |
1171 | { | |
1172 | if (PageHWPoison(page)) | |
1173 | set_page_private(page, 0); | |
1174 | } | |
1175 | ||
25182f05 NH |
1176 | /* |
1177 | * Return true if a page type of a given page is supported by hwpoison | |
1178 | * mechanism (while handling could fail), otherwise false. This function | |
1179 | * does not return true for hugetlb or device memory pages, so it's assumed | |
1180 | * to be called only in the context where we never have such pages. | |
1181 | */ | |
bf6445bc | 1182 | static inline bool HWPoisonHandlable(struct page *page, unsigned long flags) |
25182f05 | 1183 | { |
3f871370 | 1184 | /* Soft offline could migrate non-LRU movable pages */ |
bf6445bc | 1185 | if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page)) |
3f871370 | 1186 | return true; |
bf6445bc | 1187 | |
3f871370 | 1188 | return PageLRU(page) || is_free_buddy_page(page); |
25182f05 NH |
1189 | } |
1190 | ||
bf6445bc | 1191 | static int __get_hwpoison_page(struct page *page, unsigned long flags) |
ead07f6a NH |
1192 | { |
1193 | struct page *head = compound_head(page); | |
25182f05 NH |
1194 | int ret = 0; |
1195 | bool hugetlb = false; | |
1196 | ||
1197 | ret = get_hwpoison_huge_page(head, &hugetlb); | |
1198 | if (hugetlb) | |
1199 | return ret; | |
1200 | ||
1201 | /* | |
1202 | * This check prevents from calling get_hwpoison_unless_zero() | |
1203 | * for any unsupported type of page in order to reduce the risk of | |
1204 | * unexpected races caused by taking a page refcount. | |
1205 | */ | |
bf6445bc | 1206 | if (!HWPoisonHandlable(head, flags)) |
fcc00621 | 1207 | return -EBUSY; |
ead07f6a | 1208 | |
c2e7e00b KK |
1209 | if (get_page_unless_zero(head)) { |
1210 | if (head == compound_head(page)) | |
1211 | return 1; | |
1212 | ||
495367c0 CY |
1213 | pr_info("Memory failure: %#lx cannot catch tail\n", |
1214 | page_to_pfn(page)); | |
c2e7e00b KK |
1215 | put_page(head); |
1216 | } | |
1217 | ||
1218 | return 0; | |
ead07f6a | 1219 | } |
ead07f6a | 1220 | |
2f714160 | 1221 | static int get_any_page(struct page *p, unsigned long flags) |
17e395b6 | 1222 | { |
2f714160 OS |
1223 | int ret = 0, pass = 0; |
1224 | bool count_increased = false; | |
17e395b6 | 1225 | |
2f714160 OS |
1226 | if (flags & MF_COUNT_INCREASED) |
1227 | count_increased = true; | |
1228 | ||
1229 | try_again: | |
0ed950d1 | 1230 | if (!count_increased) { |
bf6445bc | 1231 | ret = __get_hwpoison_page(p, flags); |
0ed950d1 NH |
1232 | if (!ret) { |
1233 | if (page_count(p)) { | |
1234 | /* We raced with an allocation, retry. */ | |
1235 | if (pass++ < 3) | |
1236 | goto try_again; | |
1237 | ret = -EBUSY; | |
1238 | } else if (!PageHuge(p) && !is_free_buddy_page(p)) { | |
1239 | /* We raced with put_page, retry. */ | |
1240 | if (pass++ < 3) | |
1241 | goto try_again; | |
1242 | ret = -EIO; | |
1243 | } | |
1244 | goto out; | |
1245 | } else if (ret == -EBUSY) { | |
fcc00621 NH |
1246 | /* |
1247 | * We raced with (possibly temporary) unhandlable | |
1248 | * page, retry. | |
1249 | */ | |
1250 | if (pass++ < 3) { | |
d0505e9f | 1251 | shake_page(p); |
2f714160 | 1252 | goto try_again; |
fcc00621 NH |
1253 | } |
1254 | ret = -EIO; | |
0ed950d1 | 1255 | goto out; |
2f714160 | 1256 | } |
0ed950d1 NH |
1257 | } |
1258 | ||
bf6445bc | 1259 | if (PageHuge(p) || HWPoisonHandlable(p, flags)) { |
0ed950d1 | 1260 | ret = 1; |
2f714160 | 1261 | } else { |
0ed950d1 NH |
1262 | /* |
1263 | * A page we cannot handle. Check whether we can turn | |
1264 | * it into something we can handle. | |
1265 | */ | |
1266 | if (pass++ < 3) { | |
2f714160 | 1267 | put_page(p); |
d0505e9f | 1268 | shake_page(p); |
0ed950d1 NH |
1269 | count_increased = false; |
1270 | goto try_again; | |
2f714160 | 1271 | } |
0ed950d1 NH |
1272 | put_page(p); |
1273 | ret = -EIO; | |
17e395b6 | 1274 | } |
0ed950d1 | 1275 | out: |
941ca063 | 1276 | if (ret == -EIO) |
1825b93b | 1277 | pr_err("Memory failure: %#lx: unhandlable page.\n", page_to_pfn(p)); |
941ca063 | 1278 | |
17e395b6 OS |
1279 | return ret; |
1280 | } | |
1281 | ||
bf181c58 NH |
1282 | static int __get_unpoison_page(struct page *page) |
1283 | { | |
1284 | struct page *head = compound_head(page); | |
1285 | int ret = 0; | |
1286 | bool hugetlb = false; | |
1287 | ||
1288 | ret = get_hwpoison_huge_page(head, &hugetlb); | |
1289 | if (hugetlb) | |
1290 | return ret; | |
1291 | ||
1292 | /* | |
1293 | * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison, | |
1294 | * but also isolated from buddy freelist, so need to identify the | |
1295 | * state and have to cancel both operations to unpoison. | |
1296 | */ | |
1297 | if (PageHWPoisonTakenOff(page)) | |
1298 | return -EHWPOISON; | |
1299 | ||
1300 | return get_page_unless_zero(page) ? 1 : 0; | |
1301 | } | |
1302 | ||
0ed950d1 NH |
1303 | /** |
1304 | * get_hwpoison_page() - Get refcount for memory error handling | |
1305 | * @p: Raw error page (hit by memory error) | |
1306 | * @flags: Flags controlling behavior of error handling | |
1307 | * | |
1308 | * get_hwpoison_page() takes a page refcount of an error page to handle memory | |
1309 | * error on it, after checking that the error page is in a well-defined state | |
0b8f0d87 | 1310 | * (defined as a page-type we can successfully handle the memory error on it, |
0ed950d1 NH |
1311 | * such as LRU page and hugetlb page). |
1312 | * | |
1313 | * Memory error handling could be triggered at any time on any type of page, | |
1314 | * so it's prone to race with typical memory management lifecycle (like | |
1315 | * allocation and free). So to avoid such races, get_hwpoison_page() takes | |
1316 | * extra care for the error page's state (as done in __get_hwpoison_page()), | |
1317 | * and has some retry logic in get_any_page(). | |
1318 | * | |
bf181c58 NH |
1319 | * When called from unpoison_memory(), the caller should already ensure that |
1320 | * the given page has PG_hwpoison. So it's never reused for other page | |
1321 | * allocations, and __get_unpoison_page() never races with them. | |
1322 | * | |
0ed950d1 NH |
1323 | * Return: 0 on failure, |
1324 | * 1 on success for in-use pages in a well-defined state, | |
1325 | * -EIO for pages on which we can not handle memory errors, | |
1326 | * -EBUSY when get_hwpoison_page() has raced with page lifecycle | |
bf181c58 NH |
1327 | * operations like allocation and free, |
1328 | * -EHWPOISON when the page is hwpoisoned and taken off from buddy. | |
0ed950d1 NH |
1329 | */ |
1330 | static int get_hwpoison_page(struct page *p, unsigned long flags) | |
2f714160 OS |
1331 | { |
1332 | int ret; | |
1333 | ||
1334 | zone_pcp_disable(page_zone(p)); | |
bf181c58 NH |
1335 | if (flags & MF_UNPOISON) |
1336 | ret = __get_unpoison_page(p); | |
1337 | else | |
1338 | ret = get_any_page(p, flags); | |
2f714160 OS |
1339 | zone_pcp_enable(page_zone(p)); |
1340 | ||
1341 | return ret; | |
1342 | } | |
1343 | ||
6a46079c AK |
1344 | /* |
1345 | * Do all that is necessary to remove user space mappings. Unmap | |
1346 | * the pages and send SIGBUS to the processes if the data was dirty. | |
1347 | */ | |
666e5a40 | 1348 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
ed8c2f49 | 1349 | int flags, struct page *hpage) |
6a46079c | 1350 | { |
869f7ee6 | 1351 | struct folio *folio = page_folio(hpage); |
36af6737 | 1352 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC; |
6a46079c AK |
1353 | struct address_space *mapping; |
1354 | LIST_HEAD(tokill); | |
1fb08ac6 | 1355 | bool unmap_success; |
6751ed65 | 1356 | int kill = 1, forcekill; |
286c469a | 1357 | bool mlocked = PageMlocked(hpage); |
6a46079c | 1358 | |
93a9eb39 NH |
1359 | /* |
1360 | * Here we are interested only in user-mapped pages, so skip any | |
1361 | * other types of pages. | |
1362 | */ | |
1363 | if (PageReserved(p) || PageSlab(p)) | |
666e5a40 | 1364 | return true; |
93a9eb39 | 1365 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 1366 | return true; |
6a46079c | 1367 | |
6a46079c AK |
1368 | /* |
1369 | * This check implies we don't kill processes if their pages | |
1370 | * are in the swap cache early. Those are always late kills. | |
1371 | */ | |
7af446a8 | 1372 | if (!page_mapped(hpage)) |
666e5a40 | 1373 | return true; |
1668bfd5 | 1374 | |
52089b14 | 1375 | if (PageKsm(p)) { |
495367c0 | 1376 | pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn); |
666e5a40 | 1377 | return false; |
52089b14 | 1378 | } |
6a46079c AK |
1379 | |
1380 | if (PageSwapCache(p)) { | |
495367c0 CY |
1381 | pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n", |
1382 | pfn); | |
6a46079c AK |
1383 | ttu |= TTU_IGNORE_HWPOISON; |
1384 | } | |
1385 | ||
1386 | /* | |
1387 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1388 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1389 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1390 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1391 | */ |
7af446a8 | 1392 | mapping = page_mapping(hpage); |
6751ed65 | 1393 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
f56753ac | 1394 | mapping_can_writeback(mapping)) { |
7af446a8 NH |
1395 | if (page_mkclean(hpage)) { |
1396 | SetPageDirty(hpage); | |
6a46079c AK |
1397 | } else { |
1398 | kill = 0; | |
1399 | ttu |= TTU_IGNORE_HWPOISON; | |
495367c0 | 1400 | pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1401 | pfn); |
1402 | } | |
1403 | } | |
1404 | ||
1405 | /* | |
1406 | * First collect all the processes that have the page | |
1407 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1408 | * because ttu takes the rmap data structures down. | |
1409 | * | |
1410 | * Error handling: We ignore errors here because | |
1411 | * there's nothing that can be done. | |
1412 | */ | |
1413 | if (kill) | |
415c64c1 | 1414 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1415 | |
357670f7 ML |
1416 | if (PageHuge(hpage) && !PageAnon(hpage)) { |
1417 | /* | |
1418 | * For hugetlb pages in shared mappings, try_to_unmap | |
1419 | * could potentially call huge_pmd_unshare. Because of | |
1420 | * this, take semaphore in write mode here and set | |
1421 | * TTU_RMAP_LOCKED to indicate we have taken the lock | |
1422 | * at this higher level. | |
1423 | */ | |
1424 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1425 | if (mapping) { | |
9030fb0b | 1426 | try_to_unmap(folio, ttu|TTU_RMAP_LOCKED); |
357670f7 ML |
1427 | i_mmap_unlock_write(mapping); |
1428 | } else | |
1429 | pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn); | |
c0d0381a | 1430 | } else { |
9030fb0b | 1431 | try_to_unmap(folio, ttu); |
c0d0381a | 1432 | } |
1fb08ac6 YS |
1433 | |
1434 | unmap_success = !page_mapped(hpage); | |
666e5a40 | 1435 | if (!unmap_success) |
495367c0 | 1436 | pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1437 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1438 | |
286c469a NH |
1439 | /* |
1440 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1441 | * shake_page() again to ensure that it's flushed. | |
1442 | */ | |
1443 | if (mlocked) | |
d0505e9f | 1444 | shake_page(hpage); |
286c469a | 1445 | |
6a46079c AK |
1446 | /* |
1447 | * Now that the dirty bit has been propagated to the | |
1448 | * struct page and all unmaps done we can decide if | |
1449 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1450 | * was dirty or the process is not restartable, |
1451 | * otherwise the tokill list is merely | |
6a46079c AK |
1452 | * freed. When there was a problem unmapping earlier |
1453 | * use a more force-full uncatchable kill to prevent | |
1454 | * any accesses to the poisoned memory. | |
1455 | */ | |
415c64c1 | 1456 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
ae1139ec | 1457 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1458 | |
666e5a40 | 1459 | return unmap_success; |
6a46079c AK |
1460 | } |
1461 | ||
0348d2eb NH |
1462 | static int identify_page_state(unsigned long pfn, struct page *p, |
1463 | unsigned long page_flags) | |
761ad8d7 NH |
1464 | { |
1465 | struct page_state *ps; | |
0348d2eb NH |
1466 | |
1467 | /* | |
1468 | * The first check uses the current page flags which may not have any | |
1469 | * relevant information. The second check with the saved page flags is | |
1470 | * carried out only if the first check can't determine the page status. | |
1471 | */ | |
1472 | for (ps = error_states;; ps++) | |
1473 | if ((p->flags & ps->mask) == ps->res) | |
1474 | break; | |
1475 | ||
1476 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1477 | ||
1478 | if (!ps->mask) | |
1479 | for (ps = error_states;; ps++) | |
1480 | if ((page_flags & ps->mask) == ps->res) | |
1481 | break; | |
1482 | return page_action(ps, p, pfn); | |
1483 | } | |
1484 | ||
694bf0b0 OS |
1485 | static int try_to_split_thp_page(struct page *page, const char *msg) |
1486 | { | |
1487 | lock_page(page); | |
4966455d | 1488 | if (unlikely(split_huge_page(page))) { |
694bf0b0 OS |
1489 | unsigned long pfn = page_to_pfn(page); |
1490 | ||
1491 | unlock_page(page); | |
4966455d | 1492 | pr_info("%s: %#lx: thp split failed\n", msg, pfn); |
694bf0b0 OS |
1493 | put_page(page); |
1494 | return -EBUSY; | |
1495 | } | |
1496 | unlock_page(page); | |
1497 | ||
1498 | return 0; | |
1499 | } | |
1500 | ||
405ce051 NH |
1501 | /* |
1502 | * Called from hugetlb code with hugetlb_lock held. | |
1503 | * | |
1504 | * Return values: | |
1505 | * 0 - free hugepage | |
1506 | * 1 - in-use hugepage | |
1507 | * 2 - not a hugepage | |
1508 | * -EBUSY - the hugepage is busy (try to retry) | |
1509 | * -EHWPOISON - the hugepage is already hwpoisoned | |
1510 | */ | |
1511 | int __get_huge_page_for_hwpoison(unsigned long pfn, int flags) | |
1512 | { | |
1513 | struct page *page = pfn_to_page(pfn); | |
1514 | struct page *head = compound_head(page); | |
1515 | int ret = 2; /* fallback to normal page handling */ | |
1516 | bool count_increased = false; | |
1517 | ||
1518 | if (!PageHeadHuge(head)) | |
1519 | goto out; | |
1520 | ||
1521 | if (flags & MF_COUNT_INCREASED) { | |
1522 | ret = 1; | |
1523 | count_increased = true; | |
b283d983 NH |
1524 | } else if (HPageFreed(head)) { |
1525 | ret = 0; | |
1526 | } else if (HPageMigratable(head)) { | |
405ce051 NH |
1527 | ret = get_page_unless_zero(head); |
1528 | if (ret) | |
1529 | count_increased = true; | |
1530 | } else { | |
1531 | ret = -EBUSY; | |
1532 | goto out; | |
1533 | } | |
1534 | ||
1535 | if (TestSetPageHWPoison(head)) { | |
1536 | ret = -EHWPOISON; | |
1537 | goto out; | |
1538 | } | |
1539 | ||
1540 | return ret; | |
1541 | out: | |
1542 | if (count_increased) | |
1543 | put_page(head); | |
1544 | return ret; | |
1545 | } | |
1546 | ||
1547 | #ifdef CONFIG_HUGETLB_PAGE | |
1548 | /* | |
1549 | * Taking refcount of hugetlb pages needs extra care about race conditions | |
1550 | * with basic operations like hugepage allocation/free/demotion. | |
1551 | * So some of prechecks for hwpoison (pinning, and testing/setting | |
1552 | * PageHWPoison) should be done in single hugetlb_lock range. | |
1553 | */ | |
1554 | static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) | |
0348d2eb | 1555 | { |
761ad8d7 | 1556 | int res; |
405ce051 NH |
1557 | struct page *p = pfn_to_page(pfn); |
1558 | struct page *head; | |
761ad8d7 | 1559 | unsigned long page_flags; |
405ce051 | 1560 | bool retry = true; |
761ad8d7 | 1561 | |
405ce051 NH |
1562 | *hugetlb = 1; |
1563 | retry: | |
1564 | res = get_huge_page_for_hwpoison(pfn, flags); | |
1565 | if (res == 2) { /* fallback to normal page handling */ | |
1566 | *hugetlb = 0; | |
1567 | return 0; | |
1568 | } else if (res == -EHWPOISON) { | |
1569 | pr_err("Memory failure: %#lx: already hardware poisoned\n", pfn); | |
1570 | if (flags & MF_ACTION_REQUIRED) { | |
1571 | head = compound_head(p); | |
a3f5d80e | 1572 | res = kill_accessing_process(current, page_to_pfn(head), flags); |
405ce051 NH |
1573 | } |
1574 | return res; | |
1575 | } else if (res == -EBUSY) { | |
1576 | if (retry) { | |
1577 | retry = false; | |
1578 | goto retry; | |
1579 | } | |
1580 | action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); | |
a3f5d80e | 1581 | return res; |
761ad8d7 NH |
1582 | } |
1583 | ||
405ce051 NH |
1584 | head = compound_head(p); |
1585 | lock_page(head); | |
1586 | ||
1587 | if (hwpoison_filter(p)) { | |
1588 | ClearPageHWPoison(head); | |
1589 | res = -EOPNOTSUPP; | |
1590 | goto out; | |
1591 | } | |
1592 | ||
405ce051 NH |
1593 | /* |
1594 | * Handling free hugepage. The possible race with hugepage allocation | |
1595 | * or demotion can be prevented by PageHWPoison flag. | |
1596 | */ | |
1597 | if (res == 0) { | |
1598 | unlock_page(head); | |
1599 | res = MF_FAILED; | |
1600 | if (__page_handle_poison(p)) { | |
1601 | page_ref_inc(p); | |
1602 | res = MF_RECOVERED; | |
761ad8d7 | 1603 | } |
405ce051 NH |
1604 | action_result(pfn, MF_MSG_FREE_HUGE, res); |
1605 | return res == MF_RECOVERED ? 0 : -EBUSY; | |
761ad8d7 NH |
1606 | } |
1607 | ||
761ad8d7 NH |
1608 | page_flags = head->flags; |
1609 | ||
31286a84 NH |
1610 | /* |
1611 | * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so | |
1612 | * simply disable it. In order to make it work properly, we need | |
1613 | * make sure that: | |
1614 | * - conversion of a pud that maps an error hugetlb into hwpoison | |
1615 | * entry properly works, and | |
1616 | * - other mm code walking over page table is aware of pud-aligned | |
1617 | * hwpoison entries. | |
1618 | */ | |
1619 | if (huge_page_size(page_hstate(head)) > PMD_SIZE) { | |
1620 | action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED); | |
1621 | res = -EBUSY; | |
1622 | goto out; | |
1623 | } | |
1624 | ||
ed8c2f49 | 1625 | if (!hwpoison_user_mappings(p, pfn, flags, head)) { |
761ad8d7 NH |
1626 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1627 | res = -EBUSY; | |
1628 | goto out; | |
1629 | } | |
1630 | ||
ea6d0630 | 1631 | return identify_page_state(pfn, p, page_flags); |
761ad8d7 NH |
1632 | out: |
1633 | unlock_page(head); | |
1634 | return res; | |
1635 | } | |
405ce051 NH |
1636 | #else |
1637 | static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) | |
1638 | { | |
1639 | return 0; | |
1640 | } | |
1641 | #endif | |
761ad8d7 | 1642 | |
6100e34b DW |
1643 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
1644 | struct dev_pagemap *pgmap) | |
1645 | { | |
1646 | struct page *page = pfn_to_page(pfn); | |
6100e34b DW |
1647 | unsigned long size = 0; |
1648 | struct to_kill *tk; | |
1649 | LIST_HEAD(tokill); | |
1650 | int rc = -EBUSY; | |
1651 | loff_t start; | |
27359fd6 | 1652 | dax_entry_t cookie; |
6100e34b | 1653 | |
1e8aaedb OS |
1654 | if (flags & MF_COUNT_INCREASED) |
1655 | /* | |
1656 | * Drop the extra refcount in case we come from madvise(). | |
1657 | */ | |
1658 | put_page(page); | |
1659 | ||
34dc45be DW |
1660 | /* device metadata space is not recoverable */ |
1661 | if (!pgmap_pfn_valid(pgmap, pfn)) { | |
1662 | rc = -ENXIO; | |
1663 | goto out; | |
1664 | } | |
1665 | ||
61e28cf0 JM |
1666 | /* |
1667 | * Pages instantiated by device-dax (not filesystem-dax) | |
1668 | * may be compound pages. | |
1669 | */ | |
1670 | page = compound_head(page); | |
1671 | ||
6100e34b DW |
1672 | /* |
1673 | * Prevent the inode from being freed while we are interrogating | |
1674 | * the address_space, typically this would be handled by | |
1675 | * lock_page(), but dax pages do not use the page lock. This | |
1676 | * also prevents changes to the mapping of this pfn until | |
1677 | * poison signaling is complete. | |
1678 | */ | |
27359fd6 MW |
1679 | cookie = dax_lock_page(page); |
1680 | if (!cookie) | |
6100e34b DW |
1681 | goto out; |
1682 | ||
1683 | if (hwpoison_filter(page)) { | |
d1fe111f | 1684 | rc = -EOPNOTSUPP; |
6100e34b DW |
1685 | goto unlock; |
1686 | } | |
1687 | ||
25b2995a | 1688 | if (pgmap->type == MEMORY_DEVICE_PRIVATE) { |
6100e34b DW |
1689 | /* |
1690 | * TODO: Handle HMM pages which may need coordination | |
1691 | * with device-side memory. | |
1692 | */ | |
1693 | goto unlock; | |
6100e34b DW |
1694 | } |
1695 | ||
1696 | /* | |
1697 | * Use this flag as an indication that the dax page has been | |
1698 | * remapped UC to prevent speculative consumption of poison. | |
1699 | */ | |
1700 | SetPageHWPoison(page); | |
1701 | ||
1702 | /* | |
1703 | * Unlike System-RAM there is no possibility to swap in a | |
1704 | * different physical page at a given virtual address, so all | |
1705 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1706 | * SIGBUS (i.e. MF_MUST_KILL) | |
1707 | */ | |
1708 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
577553f4 | 1709 | collect_procs(page, &tokill, true); |
6100e34b DW |
1710 | |
1711 | list_for_each_entry(tk, &tokill, nd) | |
1712 | if (tk->size_shift) | |
1713 | size = max(size, 1UL << tk->size_shift); | |
1714 | if (size) { | |
1715 | /* | |
1716 | * Unmap the largest mapping to avoid breaking up | |
1717 | * device-dax mappings which are constant size. The | |
1718 | * actual size of the mapping being torn down is | |
1719 | * communicated in siginfo, see kill_proc() | |
1720 | */ | |
1721 | start = (page->index << PAGE_SHIFT) & ~(size - 1); | |
4d75136b | 1722 | unmap_mapping_range(page->mapping, start, size, 0); |
6100e34b | 1723 | } |
577553f4 | 1724 | kill_procs(&tokill, true, false, pfn, flags); |
6100e34b DW |
1725 | rc = 0; |
1726 | unlock: | |
27359fd6 | 1727 | dax_unlock_page(page, cookie); |
6100e34b DW |
1728 | out: |
1729 | /* drop pgmap ref acquired in caller */ | |
1730 | put_dev_pagemap(pgmap); | |
1731 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
1732 | return rc; | |
1733 | } | |
1734 | ||
91d00547 NH |
1735 | static DEFINE_MUTEX(mf_mutex); |
1736 | ||
cd42f4a3 TL |
1737 | /** |
1738 | * memory_failure - Handle memory failure of a page. | |
1739 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
1740 | * @flags: fine tune action taken |
1741 | * | |
1742 | * This function is called by the low level machine check code | |
1743 | * of an architecture when it detects hardware memory corruption | |
1744 | * of a page. It tries its best to recover, which includes | |
1745 | * dropping pages, killing processes etc. | |
1746 | * | |
1747 | * The function is primarily of use for corruptions that | |
1748 | * happen outside the current execution context (e.g. when | |
1749 | * detected by a background scrubber) | |
1750 | * | |
1751 | * Must run in process context (e.g. a work queue) with interrupts | |
1752 | * enabled and no spinlocks hold. | |
d1fe111f | 1753 | * |
1754 | * Return: 0 for successfully handled the memory error, | |
9113eaf3 | 1755 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event, |
d1fe111f | 1756 | * < 0(except -EOPNOTSUPP) on failure. |
cd42f4a3 | 1757 | */ |
83b57531 | 1758 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 1759 | { |
6a46079c | 1760 | struct page *p; |
7af446a8 | 1761 | struct page *hpage; |
6100e34b | 1762 | struct dev_pagemap *pgmap; |
171936dd | 1763 | int res = 0; |
524fca1e | 1764 | unsigned long page_flags; |
a8b2c2ce | 1765 | bool retry = true; |
405ce051 | 1766 | int hugetlb = 0; |
6a46079c AK |
1767 | |
1768 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 1769 | panic("Memory failure on page %lx", pfn); |
6a46079c | 1770 | |
03b122da TL |
1771 | mutex_lock(&mf_mutex); |
1772 | ||
67f22ba7 | 1773 | if (!(flags & MF_SW_SIMULATED)) |
1774 | hw_memory_failure = true; | |
1775 | ||
96c804a6 DH |
1776 | p = pfn_to_online_page(pfn); |
1777 | if (!p) { | |
03b122da TL |
1778 | res = arch_memory_failure(pfn, flags); |
1779 | if (res == 0) | |
1780 | goto unlock_mutex; | |
1781 | ||
96c804a6 DH |
1782 | if (pfn_valid(pfn)) { |
1783 | pgmap = get_dev_pagemap(pfn, NULL); | |
03b122da TL |
1784 | if (pgmap) { |
1785 | res = memory_failure_dev_pagemap(pfn, flags, | |
1786 | pgmap); | |
1787 | goto unlock_mutex; | |
1788 | } | |
96c804a6 | 1789 | } |
495367c0 CY |
1790 | pr_err("Memory failure: %#lx: memory outside kernel control\n", |
1791 | pfn); | |
03b122da TL |
1792 | res = -ENXIO; |
1793 | goto unlock_mutex; | |
6a46079c AK |
1794 | } |
1795 | ||
a8b2c2ce | 1796 | try_again: |
405ce051 NH |
1797 | res = try_memory_failure_hugetlb(pfn, flags, &hugetlb); |
1798 | if (hugetlb) | |
171936dd | 1799 | goto unlock_mutex; |
171936dd | 1800 | |
6a46079c | 1801 | if (TestSetPageHWPoison(p)) { |
495367c0 CY |
1802 | pr_err("Memory failure: %#lx: already hardware poisoned\n", |
1803 | pfn); | |
47af12ba | 1804 | res = -EHWPOISON; |
a3f5d80e NH |
1805 | if (flags & MF_ACTION_REQUIRED) |
1806 | res = kill_accessing_process(current, pfn, flags); | |
f361e246 NH |
1807 | if (flags & MF_COUNT_INCREASED) |
1808 | put_page(p); | |
171936dd | 1809 | goto unlock_mutex; |
6a46079c AK |
1810 | } |
1811 | ||
75ee64b3 | 1812 | hpage = compound_head(p); |
6a46079c AK |
1813 | |
1814 | /* | |
1815 | * We need/can do nothing about count=0 pages. | |
1816 | * 1) it's a free page, and therefore in safe hand: | |
1817 | * prep_new_page() will be the gate keeper. | |
761ad8d7 | 1818 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
1819 | * Implies some kernel user: cannot stop them from |
1820 | * R/W the page; let's pray that the page has been | |
1821 | * used and will be freed some time later. | |
1822 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 1823 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 1824 | */ |
0ed950d1 NH |
1825 | if (!(flags & MF_COUNT_INCREASED)) { |
1826 | res = get_hwpoison_page(p, flags); | |
1827 | if (!res) { | |
1828 | if (is_free_buddy_page(p)) { | |
1829 | if (take_page_off_buddy(p)) { | |
1830 | page_ref_inc(p); | |
1831 | res = MF_RECOVERED; | |
1832 | } else { | |
1833 | /* We lost the race, try again */ | |
1834 | if (retry) { | |
1835 | ClearPageHWPoison(p); | |
0ed950d1 NH |
1836 | retry = false; |
1837 | goto try_again; | |
1838 | } | |
1839 | res = MF_FAILED; | |
a8b2c2ce | 1840 | } |
0ed950d1 NH |
1841 | action_result(pfn, MF_MSG_BUDDY, res); |
1842 | res = res == MF_RECOVERED ? 0 : -EBUSY; | |
1843 | } else { | |
1844 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); | |
1845 | res = -EBUSY; | |
a8b2c2ce | 1846 | } |
0ed950d1 NH |
1847 | goto unlock_mutex; |
1848 | } else if (res < 0) { | |
1849 | action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); | |
171936dd | 1850 | res = -EBUSY; |
0ed950d1 | 1851 | goto unlock_mutex; |
8d22ba1b | 1852 | } |
6a46079c AK |
1853 | } |
1854 | ||
761ad8d7 | 1855 | if (PageTransHuge(hpage)) { |
eac96c3e YS |
1856 | /* |
1857 | * The flag must be set after the refcount is bumped | |
1858 | * otherwise it may race with THP split. | |
1859 | * And the flag can't be set in get_hwpoison_page() since | |
1860 | * it is called by soft offline too and it is just called | |
1861 | * for !MF_COUNT_INCREASE. So here seems to be the best | |
1862 | * place. | |
1863 | * | |
1864 | * Don't need care about the above error handling paths for | |
1865 | * get_hwpoison_page() since they handle either free page | |
1866 | * or unhandlable page. The refcount is bumped iff the | |
1867 | * page is a valid handlable page. | |
1868 | */ | |
1869 | SetPageHasHWPoisoned(hpage); | |
5d1fd5dc NH |
1870 | if (try_to_split_thp_page(p, "Memory Failure") < 0) { |
1871 | action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED); | |
171936dd TL |
1872 | res = -EBUSY; |
1873 | goto unlock_mutex; | |
5d1fd5dc | 1874 | } |
415c64c1 | 1875 | VM_BUG_ON_PAGE(!page_count(p), p); |
415c64c1 NH |
1876 | } |
1877 | ||
e43c3afb WF |
1878 | /* |
1879 | * We ignore non-LRU pages for good reasons. | |
1880 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 1881 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
1882 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
1883 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1884 | * walked by the page reclaim code, however that's not a big loss. | |
1885 | */ | |
d0505e9f | 1886 | shake_page(p); |
e43c3afb | 1887 | |
761ad8d7 | 1888 | lock_page(p); |
847ce401 | 1889 | |
f37d4298 | 1890 | /* |
75ee64b3 ML |
1891 | * We're only intended to deal with the non-Compound page here. |
1892 | * However, the page could have changed compound pages due to | |
1893 | * race window. If this happens, we could try again to hopefully | |
1894 | * handle the page next round. | |
f37d4298 | 1895 | */ |
75ee64b3 ML |
1896 | if (PageCompound(p)) { |
1897 | if (retry) { | |
e240ac52 | 1898 | ClearPageHWPoison(p); |
75ee64b3 ML |
1899 | unlock_page(p); |
1900 | put_page(p); | |
1901 | flags &= ~MF_COUNT_INCREASED; | |
1902 | retry = false; | |
1903 | goto try_again; | |
1904 | } | |
cc637b17 | 1905 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 | 1906 | res = -EBUSY; |
171936dd | 1907 | goto unlock_page; |
f37d4298 AK |
1908 | } |
1909 | ||
524fca1e NH |
1910 | /* |
1911 | * We use page flags to determine what action should be taken, but | |
1912 | * the flags can be modified by the error containment action. One | |
1913 | * example is an mlocked page, where PG_mlocked is cleared by | |
1914 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1915 | * correctly, we save a copy of the page flags at this time. | |
1916 | */ | |
7d9d46ac | 1917 | page_flags = p->flags; |
524fca1e | 1918 | |
7c116f2b | 1919 | if (hwpoison_filter(p)) { |
e240ac52 | 1920 | TestClearPageHWPoison(p); |
761ad8d7 | 1921 | unlock_page(p); |
dd6e2402 | 1922 | put_page(p); |
d1fe111f | 1923 | res = -EOPNOTSUPP; |
171936dd | 1924 | goto unlock_mutex; |
7c116f2b | 1925 | } |
847ce401 | 1926 | |
e8675d29 | 1927 | /* |
1928 | * __munlock_pagevec may clear a writeback page's LRU flag without | |
1929 | * page_lock. We need wait writeback completion for this page or it | |
1930 | * may trigger vfs BUG while evict inode. | |
1931 | */ | |
b04d3eeb | 1932 | if (!PageLRU(p) && !PageWriteback(p)) |
0bc1f8b0 CY |
1933 | goto identify_page_state; |
1934 | ||
6edd6cc6 NH |
1935 | /* |
1936 | * It's very difficult to mess with pages currently under IO | |
1937 | * and in many cases impossible, so we just avoid it here. | |
1938 | */ | |
6a46079c AK |
1939 | wait_on_page_writeback(p); |
1940 | ||
1941 | /* | |
1942 | * Now take care of user space mappings. | |
e64a782f | 1943 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
6a46079c | 1944 | */ |
ed8c2f49 | 1945 | if (!hwpoison_user_mappings(p, pfn, flags, p)) { |
cc637b17 | 1946 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 | 1947 | res = -EBUSY; |
171936dd | 1948 | goto unlock_page; |
1668bfd5 | 1949 | } |
6a46079c AK |
1950 | |
1951 | /* | |
1952 | * Torn down by someone else? | |
1953 | */ | |
dc2a1cbf | 1954 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 1955 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 1956 | res = -EBUSY; |
171936dd | 1957 | goto unlock_page; |
6a46079c AK |
1958 | } |
1959 | ||
0bc1f8b0 | 1960 | identify_page_state: |
0348d2eb | 1961 | res = identify_page_state(pfn, p, page_flags); |
ea6d0630 NH |
1962 | mutex_unlock(&mf_mutex); |
1963 | return res; | |
171936dd | 1964 | unlock_page: |
761ad8d7 | 1965 | unlock_page(p); |
171936dd TL |
1966 | unlock_mutex: |
1967 | mutex_unlock(&mf_mutex); | |
6a46079c AK |
1968 | return res; |
1969 | } | |
cd42f4a3 | 1970 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1971 | |
ea8f5fb8 HY |
1972 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1973 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1974 | ||
1975 | struct memory_failure_entry { | |
1976 | unsigned long pfn; | |
ea8f5fb8 HY |
1977 | int flags; |
1978 | }; | |
1979 | ||
1980 | struct memory_failure_cpu { | |
1981 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1982 | MEMORY_FAILURE_FIFO_SIZE); | |
1983 | spinlock_t lock; | |
1984 | struct work_struct work; | |
1985 | }; | |
1986 | ||
1987 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1988 | ||
1989 | /** | |
1990 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1991 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
1992 | * @flags: Flags for memory failure handling |
1993 | * | |
1994 | * This function is called by the low level hardware error handler | |
1995 | * when it detects hardware memory corruption of a page. It schedules | |
1996 | * the recovering of error page, including dropping pages, killing | |
1997 | * processes etc. | |
1998 | * | |
1999 | * The function is primarily of use for corruptions that | |
2000 | * happen outside the current execution context (e.g. when | |
2001 | * detected by a background scrubber) | |
2002 | * | |
2003 | * Can run in IRQ context. | |
2004 | */ | |
83b57531 | 2005 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
2006 | { |
2007 | struct memory_failure_cpu *mf_cpu; | |
2008 | unsigned long proc_flags; | |
2009 | struct memory_failure_entry entry = { | |
2010 | .pfn = pfn, | |
ea8f5fb8 HY |
2011 | .flags = flags, |
2012 | }; | |
2013 | ||
2014 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
2015 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 2016 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
2017 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
2018 | else | |
8e33a52f | 2019 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
2020 | pfn); |
2021 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
2022 | put_cpu_var(memory_failure_cpu); | |
2023 | } | |
2024 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
2025 | ||
2026 | static void memory_failure_work_func(struct work_struct *work) | |
2027 | { | |
2028 | struct memory_failure_cpu *mf_cpu; | |
2029 | struct memory_failure_entry entry = { 0, }; | |
2030 | unsigned long proc_flags; | |
2031 | int gotten; | |
2032 | ||
06202231 | 2033 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
ea8f5fb8 HY |
2034 | for (;;) { |
2035 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
2036 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
2037 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
2038 | if (!gotten) | |
2039 | break; | |
cf870c70 | 2040 | if (entry.flags & MF_SOFT_OFFLINE) |
feec24a6 | 2041 | soft_offline_page(entry.pfn, entry.flags); |
cf870c70 | 2042 | else |
83b57531 | 2043 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
2044 | } |
2045 | } | |
2046 | ||
06202231 JM |
2047 | /* |
2048 | * Process memory_failure work queued on the specified CPU. | |
2049 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. | |
2050 | */ | |
2051 | void memory_failure_queue_kick(int cpu) | |
2052 | { | |
2053 | struct memory_failure_cpu *mf_cpu; | |
2054 | ||
2055 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
2056 | cancel_work_sync(&mf_cpu->work); | |
2057 | memory_failure_work_func(&mf_cpu->work); | |
2058 | } | |
2059 | ||
ea8f5fb8 HY |
2060 | static int __init memory_failure_init(void) |
2061 | { | |
2062 | struct memory_failure_cpu *mf_cpu; | |
2063 | int cpu; | |
2064 | ||
2065 | for_each_possible_cpu(cpu) { | |
2066 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
2067 | spin_lock_init(&mf_cpu->lock); | |
2068 | INIT_KFIFO(mf_cpu->fifo); | |
2069 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
2070 | } | |
2071 | ||
2072 | return 0; | |
2073 | } | |
2074 | core_initcall(memory_failure_init); | |
2075 | ||
a5f65109 NH |
2076 | #define unpoison_pr_info(fmt, pfn, rs) \ |
2077 | ({ \ | |
2078 | if (__ratelimit(rs)) \ | |
2079 | pr_info(fmt, pfn); \ | |
2080 | }) | |
2081 | ||
847ce401 WF |
2082 | /** |
2083 | * unpoison_memory - Unpoison a previously poisoned page | |
2084 | * @pfn: Page number of the to be unpoisoned page | |
2085 | * | |
2086 | * Software-unpoison a page that has been poisoned by | |
2087 | * memory_failure() earlier. | |
2088 | * | |
2089 | * This is only done on the software-level, so it only works | |
2090 | * for linux injected failures, not real hardware failures | |
2091 | * | |
2092 | * Returns 0 for success, otherwise -errno. | |
2093 | */ | |
2094 | int unpoison_memory(unsigned long pfn) | |
2095 | { | |
2096 | struct page *page; | |
2097 | struct page *p; | |
bf181c58 | 2098 | int ret = -EBUSY; |
c8bd84f7 | 2099 | int freeit = 0; |
a5f65109 NH |
2100 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
2101 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
2102 | |
2103 | if (!pfn_valid(pfn)) | |
2104 | return -ENXIO; | |
2105 | ||
2106 | p = pfn_to_page(pfn); | |
2107 | page = compound_head(p); | |
2108 | ||
91d00547 NH |
2109 | mutex_lock(&mf_mutex); |
2110 | ||
67f22ba7 | 2111 | if (hw_memory_failure) { |
2112 | unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n", | |
2113 | pfn, &unpoison_rs); | |
2114 | ret = -EOPNOTSUPP; | |
2115 | goto unlock_mutex; | |
2116 | } | |
2117 | ||
847ce401 | 2118 | if (!PageHWPoison(p)) { |
495367c0 | 2119 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 2120 | pfn, &unpoison_rs); |
91d00547 | 2121 | goto unlock_mutex; |
847ce401 WF |
2122 | } |
2123 | ||
230ac719 | 2124 | if (page_count(page) > 1) { |
495367c0 | 2125 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 2126 | pfn, &unpoison_rs); |
91d00547 | 2127 | goto unlock_mutex; |
230ac719 NH |
2128 | } |
2129 | ||
2130 | if (page_mapped(page)) { | |
495367c0 | 2131 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 2132 | pfn, &unpoison_rs); |
91d00547 | 2133 | goto unlock_mutex; |
230ac719 NH |
2134 | } |
2135 | ||
2136 | if (page_mapping(page)) { | |
495367c0 | 2137 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 2138 | pfn, &unpoison_rs); |
91d00547 | 2139 | goto unlock_mutex; |
0cea3fdc WL |
2140 | } |
2141 | ||
bf181c58 | 2142 | if (PageSlab(page) || PageTable(page)) |
91d00547 | 2143 | goto unlock_mutex; |
847ce401 | 2144 | |
bf181c58 NH |
2145 | ret = get_hwpoison_page(p, MF_UNPOISON); |
2146 | if (!ret) { | |
c8bd84f7 | 2147 | ret = TestClearPageHWPoison(page) ? 0 : -EBUSY; |
bf181c58 NH |
2148 | } else if (ret < 0) { |
2149 | if (ret == -EHWPOISON) { | |
c8bd84f7 | 2150 | ret = put_page_back_buddy(p) ? 0 : -EBUSY; |
bf181c58 NH |
2151 | } else |
2152 | unpoison_pr_info("Unpoison: failed to grab page %#lx\n", | |
2153 | pfn, &unpoison_rs); | |
2154 | } else { | |
c8bd84f7 | 2155 | freeit = !!TestClearPageHWPoison(p); |
847ce401 | 2156 | |
dd6e2402 | 2157 | put_page(page); |
bf181c58 NH |
2158 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) { |
2159 | put_page(page); | |
2160 | ret = 0; | |
2161 | } | |
2162 | } | |
847ce401 | 2163 | |
91d00547 NH |
2164 | unlock_mutex: |
2165 | mutex_unlock(&mf_mutex); | |
c8bd84f7 | 2166 | if (!ret || freeit) { |
2167 | num_poisoned_pages_dec(); | |
2168 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", | |
2169 | page_to_pfn(p), &unpoison_rs); | |
2170 | } | |
91d00547 | 2171 | return ret; |
847ce401 WF |
2172 | } |
2173 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 2174 | |
6b9a217e | 2175 | static bool isolate_page(struct page *page, struct list_head *pagelist) |
d950b958 | 2176 | { |
6b9a217e OS |
2177 | bool isolated = false; |
2178 | bool lru = PageLRU(page); | |
d950b958 | 2179 | |
6b9a217e | 2180 | if (PageHuge(page)) { |
7ce82f4c | 2181 | isolated = !isolate_hugetlb(page, pagelist); |
6b9a217e OS |
2182 | } else { |
2183 | if (lru) | |
2184 | isolated = !isolate_lru_page(page); | |
2185 | else | |
2186 | isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE); | |
2187 | ||
2188 | if (isolated) | |
2189 | list_add(&page->lru, pagelist); | |
0ebff32c | 2190 | } |
d950b958 | 2191 | |
6b9a217e OS |
2192 | if (isolated && lru) |
2193 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
2194 | page_is_file_lru(page)); | |
2195 | ||
03613808 | 2196 | /* |
6b9a217e OS |
2197 | * If we succeed to isolate the page, we grabbed another refcount on |
2198 | * the page, so we can safely drop the one we got from get_any_pages(). | |
2199 | * If we failed to isolate the page, it means that we cannot go further | |
2200 | * and we will return an error, so drop the reference we got from | |
2201 | * get_any_pages() as well. | |
03613808 | 2202 | */ |
6b9a217e OS |
2203 | put_page(page); |
2204 | return isolated; | |
d950b958 NH |
2205 | } |
2206 | ||
6b9a217e OS |
2207 | /* |
2208 | * __soft_offline_page handles hugetlb-pages and non-hugetlb pages. | |
2209 | * If the page is a non-dirty unmapped page-cache page, it simply invalidates. | |
2210 | * If the page is mapped, it migrates the contents over. | |
2211 | */ | |
2212 | static int __soft_offline_page(struct page *page) | |
af8fae7c | 2213 | { |
d6c75dc2 | 2214 | long ret = 0; |
af8fae7c | 2215 | unsigned long pfn = page_to_pfn(page); |
6b9a217e OS |
2216 | struct page *hpage = compound_head(page); |
2217 | char const *msg_page[] = {"page", "hugepage"}; | |
2218 | bool huge = PageHuge(page); | |
2219 | LIST_HEAD(pagelist); | |
54608759 JK |
2220 | struct migration_target_control mtc = { |
2221 | .nid = NUMA_NO_NODE, | |
2222 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
2223 | }; | |
facb6011 | 2224 | |
0ebff32c | 2225 | lock_page(page); |
6b9a217e OS |
2226 | if (!PageHuge(page)) |
2227 | wait_on_page_writeback(page); | |
af8fae7c NH |
2228 | if (PageHWPoison(page)) { |
2229 | unlock_page(page); | |
dd6e2402 | 2230 | put_page(page); |
af8fae7c | 2231 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
5a2ffca3 | 2232 | return 0; |
af8fae7c | 2233 | } |
6b9a217e | 2234 | |
593396b8 | 2235 | if (!PageHuge(page) && PageLRU(page) && !PageSwapCache(page)) |
6b9a217e OS |
2236 | /* |
2237 | * Try to invalidate first. This should work for | |
2238 | * non dirty unmapped page cache pages. | |
2239 | */ | |
2240 | ret = invalidate_inode_page(page); | |
facb6011 | 2241 | unlock_page(page); |
6b9a217e | 2242 | |
6b9a217e | 2243 | if (ret) { |
fb46e735 | 2244 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
6b9a217e | 2245 | page_handle_poison(page, false, true); |
af8fae7c | 2246 | return 0; |
facb6011 AK |
2247 | } |
2248 | ||
6b9a217e | 2249 | if (isolate_page(hpage, &pagelist)) { |
54608759 | 2250 | ret = migrate_pages(&pagelist, alloc_migration_target, NULL, |
5ac95884 | 2251 | (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL); |
79f5f8fa | 2252 | if (!ret) { |
6b9a217e OS |
2253 | bool release = !huge; |
2254 | ||
2255 | if (!page_handle_poison(page, huge, release)) | |
2256 | ret = -EBUSY; | |
79f5f8fa | 2257 | } else { |
85fbe5d1 YX |
2258 | if (!list_empty(&pagelist)) |
2259 | putback_movable_pages(&pagelist); | |
59c82b70 | 2260 | |
d6c75dc2 | 2261 | pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n", |
23efd080 | 2262 | pfn, msg_page[huge], ret, &page->flags); |
facb6011 | 2263 | if (ret > 0) |
3f4b815a | 2264 | ret = -EBUSY; |
facb6011 AK |
2265 | } |
2266 | } else { | |
23efd080 MWO |
2267 | pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n", |
2268 | pfn, msg_page[huge], page_count(page), &page->flags); | |
6b9a217e | 2269 | ret = -EBUSY; |
facb6011 | 2270 | } |
facb6011 AK |
2271 | return ret; |
2272 | } | |
86e05773 | 2273 | |
6b9a217e | 2274 | static int soft_offline_in_use_page(struct page *page) |
acc14dc4 | 2275 | { |
acc14dc4 NH |
2276 | struct page *hpage = compound_head(page); |
2277 | ||
694bf0b0 OS |
2278 | if (!PageHuge(page) && PageTransHuge(hpage)) |
2279 | if (try_to_split_thp_page(page, "soft offline") < 0) | |
acc14dc4 | 2280 | return -EBUSY; |
6b9a217e | 2281 | return __soft_offline_page(page); |
acc14dc4 NH |
2282 | } |
2283 | ||
d4ae9916 | 2284 | static int soft_offline_free_page(struct page *page) |
acc14dc4 | 2285 | { |
6b9a217e | 2286 | int rc = 0; |
acc14dc4 | 2287 | |
6b9a217e OS |
2288 | if (!page_handle_poison(page, true, false)) |
2289 | rc = -EBUSY; | |
06be6ff3 | 2290 | |
d4ae9916 | 2291 | return rc; |
acc14dc4 NH |
2292 | } |
2293 | ||
dad4e5b3 DW |
2294 | static void put_ref_page(struct page *page) |
2295 | { | |
2296 | if (page) | |
2297 | put_page(page); | |
2298 | } | |
2299 | ||
86e05773 WL |
2300 | /** |
2301 | * soft_offline_page - Soft offline a page. | |
feec24a6 | 2302 | * @pfn: pfn to soft-offline |
86e05773 WL |
2303 | * @flags: flags. Same as memory_failure(). |
2304 | * | |
9113eaf3 | 2305 | * Returns 0 on success |
2306 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event | |
2307 | * < 0 otherwise negated errno. | |
86e05773 WL |
2308 | * |
2309 | * Soft offline a page, by migration or invalidation, | |
2310 | * without killing anything. This is for the case when | |
2311 | * a page is not corrupted yet (so it's still valid to access), | |
2312 | * but has had a number of corrected errors and is better taken | |
2313 | * out. | |
2314 | * | |
2315 | * The actual policy on when to do that is maintained by | |
2316 | * user space. | |
2317 | * | |
2318 | * This should never impact any application or cause data loss, | |
2319 | * however it might take some time. | |
2320 | * | |
2321 | * This is not a 100% solution for all memory, but tries to be | |
2322 | * ``good enough'' for the majority of memory. | |
2323 | */ | |
feec24a6 | 2324 | int soft_offline_page(unsigned long pfn, int flags) |
86e05773 WL |
2325 | { |
2326 | int ret; | |
b94e0282 | 2327 | bool try_again = true; |
dad4e5b3 DW |
2328 | struct page *page, *ref_page = NULL; |
2329 | ||
2330 | WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED)); | |
86e05773 | 2331 | |
feec24a6 NH |
2332 | if (!pfn_valid(pfn)) |
2333 | return -ENXIO; | |
dad4e5b3 DW |
2334 | if (flags & MF_COUNT_INCREASED) |
2335 | ref_page = pfn_to_page(pfn); | |
2336 | ||
feec24a6 NH |
2337 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ |
2338 | page = pfn_to_online_page(pfn); | |
dad4e5b3 DW |
2339 | if (!page) { |
2340 | put_ref_page(ref_page); | |
86a66810 | 2341 | return -EIO; |
dad4e5b3 | 2342 | } |
86a66810 | 2343 | |
91d00547 NH |
2344 | mutex_lock(&mf_mutex); |
2345 | ||
86e05773 | 2346 | if (PageHWPoison(page)) { |
8295d535 | 2347 | pr_info("%s: %#lx page already poisoned\n", __func__, pfn); |
dad4e5b3 | 2348 | put_ref_page(ref_page); |
91d00547 | 2349 | mutex_unlock(&mf_mutex); |
5a2ffca3 | 2350 | return 0; |
86e05773 | 2351 | } |
86e05773 | 2352 | |
b94e0282 | 2353 | retry: |
bfc8c901 | 2354 | get_online_mems(); |
bf6445bc | 2355 | ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE); |
bfc8c901 | 2356 | put_online_mems(); |
4e41a30c | 2357 | |
9113eaf3 | 2358 | if (hwpoison_filter(page)) { |
2359 | if (ret > 0) | |
2360 | put_page(page); | |
2361 | else | |
2362 | put_ref_page(ref_page); | |
2363 | ||
2364 | mutex_unlock(&mf_mutex); | |
2365 | return -EOPNOTSUPP; | |
2366 | } | |
2367 | ||
8295d535 | 2368 | if (ret > 0) { |
6b9a217e | 2369 | ret = soft_offline_in_use_page(page); |
8295d535 | 2370 | } else if (ret == 0) { |
b94e0282 OS |
2371 | if (soft_offline_free_page(page) && try_again) { |
2372 | try_again = false; | |
2a57d83c | 2373 | flags &= ~MF_COUNT_INCREASED; |
b94e0282 OS |
2374 | goto retry; |
2375 | } | |
8295d535 | 2376 | } |
4e41a30c | 2377 | |
91d00547 NH |
2378 | mutex_unlock(&mf_mutex); |
2379 | ||
86e05773 WL |
2380 | return ret; |
2381 | } | |
60f272f6 | 2382 | |
2383 | void clear_hwpoisoned_pages(struct page *memmap, int nr_pages) | |
2384 | { | |
2385 | int i; | |
2386 | ||
2387 | /* | |
2388 | * A further optimization is to have per section refcounted | |
2389 | * num_poisoned_pages. But that would need more space per memmap, so | |
2390 | * for now just do a quick global check to speed up this routine in the | |
2391 | * absence of bad pages. | |
2392 | */ | |
2393 | if (atomic_long_read(&num_poisoned_pages) == 0) | |
2394 | return; | |
2395 | ||
2396 | for (i = 0; i < nr_pages; i++) { | |
2397 | if (PageHWPoison(&memmap[i])) { | |
2398 | num_poisoned_pages_dec(); | |
2399 | ClearPageHWPoison(&memmap[i]); | |
2400 | } | |
2401 | } | |
2402 | } |