Commit | Line | Data |
---|---|---|
1439f94c | 1 | // SPDX-License-Identifier: GPL-2.0-only |
6a46079c AK |
2 | /* |
3 | * Copyright (C) 2008, 2009 Intel Corporation | |
4 | * Authors: Andi Kleen, Fengguang Wu | |
5 | * | |
6a46079c | 6 | * High level machine check handler. Handles pages reported by the |
1c80b990 | 7 | * hardware as being corrupted usually due to a multi-bit ECC memory or cache |
6a46079c | 8 | * failure. |
1c80b990 AK |
9 | * |
10 | * In addition there is a "soft offline" entry point that allows stop using | |
11 | * not-yet-corrupted-by-suspicious pages without killing anything. | |
6a46079c AK |
12 | * |
13 | * Handles page cache pages in various states. The tricky part | |
1c80b990 AK |
14 | * here is that we can access any page asynchronously in respect to |
15 | * other VM users, because memory failures could happen anytime and | |
16 | * anywhere. This could violate some of their assumptions. This is why | |
17 | * this code has to be extremely careful. Generally it tries to use | |
18 | * normal locking rules, as in get the standard locks, even if that means | |
19 | * the error handling takes potentially a long time. | |
e0de78df AK |
20 | * |
21 | * It can be very tempting to add handling for obscure cases here. | |
22 | * In general any code for handling new cases should only be added iff: | |
23 | * - You know how to test it. | |
24 | * - You have a test that can be added to mce-test | |
25 | * https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/ | |
26 | * - The case actually shows up as a frequent (top 10) page state in | |
27 | * tools/vm/page-types when running a real workload. | |
1c80b990 AK |
28 | * |
29 | * There are several operations here with exponential complexity because | |
30 | * of unsuitable VM data structures. For example the operation to map back | |
31 | * from RMAP chains to processes has to walk the complete process list and | |
32 | * has non linear complexity with the number. But since memory corruptions | |
33 | * are rare we hope to get away with this. This avoids impacting the core | |
34 | * VM. | |
6a46079c | 35 | */ |
6a46079c AK |
36 | #include <linux/kernel.h> |
37 | #include <linux/mm.h> | |
38 | #include <linux/page-flags.h> | |
478c5ffc | 39 | #include <linux/kernel-page-flags.h> |
3f07c014 | 40 | #include <linux/sched/signal.h> |
29930025 | 41 | #include <linux/sched/task.h> |
01e00f88 | 42 | #include <linux/ksm.h> |
6a46079c | 43 | #include <linux/rmap.h> |
b9e15baf | 44 | #include <linux/export.h> |
6a46079c AK |
45 | #include <linux/pagemap.h> |
46 | #include <linux/swap.h> | |
47 | #include <linux/backing-dev.h> | |
facb6011 | 48 | #include <linux/migrate.h> |
facb6011 | 49 | #include <linux/suspend.h> |
5a0e3ad6 | 50 | #include <linux/slab.h> |
bf998156 | 51 | #include <linux/swapops.h> |
7af446a8 | 52 | #include <linux/hugetlb.h> |
20d6c96b | 53 | #include <linux/memory_hotplug.h> |
5db8a73a | 54 | #include <linux/mm_inline.h> |
6100e34b | 55 | #include <linux/memremap.h> |
ea8f5fb8 | 56 | #include <linux/kfifo.h> |
a5f65109 | 57 | #include <linux/ratelimit.h> |
d4ae9916 | 58 | #include <linux/page-isolation.h> |
a3f5d80e | 59 | #include <linux/pagewalk.h> |
6a46079c | 60 | #include "internal.h" |
97f0b134 | 61 | #include "ras/ras_event.h" |
6a46079c AK |
62 | |
63 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
64 | ||
65 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
66 | ||
293c07e3 | 67 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 68 | |
6b9a217e | 69 | static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release) |
06be6ff3 | 70 | { |
6b9a217e OS |
71 | if (hugepage_or_freepage) { |
72 | /* | |
73 | * Doing this check for free pages is also fine since dissolve_free_huge_page | |
74 | * returns 0 for non-hugetlb pages as well. | |
75 | */ | |
76 | if (dissolve_free_huge_page(page) || !take_page_off_buddy(page)) | |
77 | /* | |
78 | * We could fail to take off the target page from buddy | |
f0953a1b | 79 | * for example due to racy page allocation, but that's |
6b9a217e OS |
80 | * acceptable because soft-offlined page is not broken |
81 | * and if someone really want to use it, they should | |
82 | * take it. | |
83 | */ | |
84 | return false; | |
85 | } | |
86 | ||
06be6ff3 | 87 | SetPageHWPoison(page); |
79f5f8fa OS |
88 | if (release) |
89 | put_page(page); | |
06be6ff3 OS |
90 | page_ref_inc(page); |
91 | num_poisoned_pages_inc(); | |
6b9a217e OS |
92 | |
93 | return true; | |
06be6ff3 OS |
94 | } |
95 | ||
27df5068 AK |
96 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
97 | ||
1bfe5feb | 98 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
99 | u32 hwpoison_filter_dev_major = ~0U; |
100 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
101 | u64 hwpoison_filter_flags_mask; |
102 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 103 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
104 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
105 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
106 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
107 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
108 | |
109 | static int hwpoison_filter_dev(struct page *p) | |
110 | { | |
111 | struct address_space *mapping; | |
112 | dev_t dev; | |
113 | ||
114 | if (hwpoison_filter_dev_major == ~0U && | |
115 | hwpoison_filter_dev_minor == ~0U) | |
116 | return 0; | |
117 | ||
118 | /* | |
1c80b990 | 119 | * page_mapping() does not accept slab pages. |
7c116f2b WF |
120 | */ |
121 | if (PageSlab(p)) | |
122 | return -EINVAL; | |
123 | ||
124 | mapping = page_mapping(p); | |
125 | if (mapping == NULL || mapping->host == NULL) | |
126 | return -EINVAL; | |
127 | ||
128 | dev = mapping->host->i_sb->s_dev; | |
129 | if (hwpoison_filter_dev_major != ~0U && | |
130 | hwpoison_filter_dev_major != MAJOR(dev)) | |
131 | return -EINVAL; | |
132 | if (hwpoison_filter_dev_minor != ~0U && | |
133 | hwpoison_filter_dev_minor != MINOR(dev)) | |
134 | return -EINVAL; | |
135 | ||
136 | return 0; | |
137 | } | |
138 | ||
478c5ffc WF |
139 | static int hwpoison_filter_flags(struct page *p) |
140 | { | |
141 | if (!hwpoison_filter_flags_mask) | |
142 | return 0; | |
143 | ||
144 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
145 | hwpoison_filter_flags_value) | |
146 | return 0; | |
147 | else | |
148 | return -EINVAL; | |
149 | } | |
150 | ||
4fd466eb AK |
151 | /* |
152 | * This allows stress tests to limit test scope to a collection of tasks | |
153 | * by putting them under some memcg. This prevents killing unrelated/important | |
154 | * processes such as /sbin/init. Note that the target task may share clean | |
155 | * pages with init (eg. libc text), which is harmless. If the target task | |
156 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
157 | * is also included in the memcg. At last, due to race conditions this filter | |
158 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
159 | * a freed page. | |
160 | */ | |
94a59fb3 | 161 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
162 | u64 hwpoison_filter_memcg; |
163 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
164 | static int hwpoison_filter_task(struct page *p) | |
165 | { | |
4fd466eb AK |
166 | if (!hwpoison_filter_memcg) |
167 | return 0; | |
168 | ||
94a59fb3 | 169 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
170 | return -EINVAL; |
171 | ||
172 | return 0; | |
173 | } | |
174 | #else | |
175 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
176 | #endif | |
177 | ||
7c116f2b WF |
178 | int hwpoison_filter(struct page *p) |
179 | { | |
1bfe5feb HL |
180 | if (!hwpoison_filter_enable) |
181 | return 0; | |
182 | ||
7c116f2b WF |
183 | if (hwpoison_filter_dev(p)) |
184 | return -EINVAL; | |
185 | ||
478c5ffc WF |
186 | if (hwpoison_filter_flags(p)) |
187 | return -EINVAL; | |
188 | ||
4fd466eb AK |
189 | if (hwpoison_filter_task(p)) |
190 | return -EINVAL; | |
191 | ||
7c116f2b WF |
192 | return 0; |
193 | } | |
27df5068 AK |
194 | #else |
195 | int hwpoison_filter(struct page *p) | |
196 | { | |
197 | return 0; | |
198 | } | |
199 | #endif | |
200 | ||
7c116f2b WF |
201 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
202 | ||
ae1139ec DW |
203 | /* |
204 | * Kill all processes that have a poisoned page mapped and then isolate | |
205 | * the page. | |
206 | * | |
207 | * General strategy: | |
208 | * Find all processes having the page mapped and kill them. | |
209 | * But we keep a page reference around so that the page is not | |
210 | * actually freed yet. | |
211 | * Then stash the page away | |
212 | * | |
213 | * There's no convenient way to get back to mapped processes | |
214 | * from the VMAs. So do a brute-force search over all | |
215 | * running processes. | |
216 | * | |
217 | * Remember that machine checks are not common (or rather | |
218 | * if they are common you have other problems), so this shouldn't | |
219 | * be a performance issue. | |
220 | * | |
221 | * Also there are some races possible while we get from the | |
222 | * error detection to actually handle it. | |
223 | */ | |
224 | ||
225 | struct to_kill { | |
226 | struct list_head nd; | |
227 | struct task_struct *tsk; | |
228 | unsigned long addr; | |
229 | short size_shift; | |
ae1139ec DW |
230 | }; |
231 | ||
6a46079c | 232 | /* |
7329bbeb TL |
233 | * Send all the processes who have the page mapped a signal. |
234 | * ``action optional'' if they are not immediately affected by the error | |
235 | * ``action required'' if error happened in current execution context | |
6a46079c | 236 | */ |
ae1139ec | 237 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 238 | { |
ae1139ec DW |
239 | struct task_struct *t = tk->tsk; |
240 | short addr_lsb = tk->size_shift; | |
872e9a20 | 241 | int ret = 0; |
6a46079c | 242 | |
03151c6e | 243 | pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", |
872e9a20 | 244 | pfn, t->comm, t->pid); |
7329bbeb | 245 | |
872e9a20 | 246 | if (flags & MF_ACTION_REQUIRED) { |
30c9cf49 AY |
247 | if (t == current) |
248 | ret = force_sig_mceerr(BUS_MCEERR_AR, | |
872e9a20 | 249 | (void __user *)tk->addr, addr_lsb); |
30c9cf49 AY |
250 | else |
251 | /* Signal other processes sharing the page if they have PF_MCE_EARLY set. */ | |
252 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, | |
253 | addr_lsb, t); | |
7329bbeb TL |
254 | } else { |
255 | /* | |
256 | * Don't use force here, it's convenient if the signal | |
257 | * can be temporarily blocked. | |
258 | * This could cause a loop when the user sets SIGBUS | |
259 | * to SIG_IGN, but hopefully no one will do that? | |
260 | */ | |
ae1139ec | 261 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
c0f45555 | 262 | addr_lsb, t); /* synchronous? */ |
7329bbeb | 263 | } |
6a46079c | 264 | if (ret < 0) |
495367c0 | 265 | pr_info("Memory failure: Error sending signal to %s:%d: %d\n", |
1170532b | 266 | t->comm, t->pid, ret); |
6a46079c AK |
267 | return ret; |
268 | } | |
269 | ||
588f9ce6 | 270 | /* |
47e431f4 OS |
271 | * Unknown page type encountered. Try to check whether it can turn PageLRU by |
272 | * lru_add_drain_all, or a free page by reclaiming slabs when possible. | |
588f9ce6 | 273 | */ |
facb6011 | 274 | void shake_page(struct page *p, int access) |
588f9ce6 | 275 | { |
8bcb74de NH |
276 | if (PageHuge(p)) |
277 | return; | |
278 | ||
588f9ce6 AK |
279 | if (!PageSlab(p)) { |
280 | lru_add_drain_all(); | |
588f9ce6 AK |
281 | if (PageLRU(p) || is_free_buddy_page(p)) |
282 | return; | |
283 | } | |
facb6011 | 284 | |
588f9ce6 | 285 | /* |
6b4f7799 JW |
286 | * Only call shrink_node_slabs here (which would also shrink |
287 | * other caches) if access is not potentially fatal. | |
588f9ce6 | 288 | */ |
cb731d6c VD |
289 | if (access) |
290 | drop_slab_node(page_to_nid(p)); | |
588f9ce6 AK |
291 | } |
292 | EXPORT_SYMBOL_GPL(shake_page); | |
293 | ||
6100e34b DW |
294 | static unsigned long dev_pagemap_mapping_shift(struct page *page, |
295 | struct vm_area_struct *vma) | |
296 | { | |
297 | unsigned long address = vma_address(page, vma); | |
298 | pgd_t *pgd; | |
299 | p4d_t *p4d; | |
300 | pud_t *pud; | |
301 | pmd_t *pmd; | |
302 | pte_t *pte; | |
303 | ||
304 | pgd = pgd_offset(vma->vm_mm, address); | |
305 | if (!pgd_present(*pgd)) | |
306 | return 0; | |
307 | p4d = p4d_offset(pgd, address); | |
308 | if (!p4d_present(*p4d)) | |
309 | return 0; | |
310 | pud = pud_offset(p4d, address); | |
311 | if (!pud_present(*pud)) | |
312 | return 0; | |
313 | if (pud_devmap(*pud)) | |
314 | return PUD_SHIFT; | |
315 | pmd = pmd_offset(pud, address); | |
316 | if (!pmd_present(*pmd)) | |
317 | return 0; | |
318 | if (pmd_devmap(*pmd)) | |
319 | return PMD_SHIFT; | |
320 | pte = pte_offset_map(pmd, address); | |
321 | if (!pte_present(*pte)) | |
322 | return 0; | |
323 | if (pte_devmap(*pte)) | |
324 | return PAGE_SHIFT; | |
325 | return 0; | |
326 | } | |
6a46079c AK |
327 | |
328 | /* | |
329 | * Failure handling: if we can't find or can't kill a process there's | |
330 | * not much we can do. We just print a message and ignore otherwise. | |
331 | */ | |
332 | ||
333 | /* | |
334 | * Schedule a process for later kill. | |
335 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
6a46079c AK |
336 | */ |
337 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
338 | struct vm_area_struct *vma, | |
996ff7a0 | 339 | struct list_head *to_kill) |
6a46079c AK |
340 | { |
341 | struct to_kill *tk; | |
342 | ||
996ff7a0 JC |
343 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
344 | if (!tk) { | |
345 | pr_err("Memory failure: Out of memory while machine check handling\n"); | |
346 | return; | |
6a46079c | 347 | } |
996ff7a0 | 348 | |
6a46079c | 349 | tk->addr = page_address_in_vma(p, vma); |
6100e34b DW |
350 | if (is_zone_device_page(p)) |
351 | tk->size_shift = dev_pagemap_mapping_shift(p, vma); | |
352 | else | |
75068518 | 353 | tk->size_shift = page_shift(compound_head(p)); |
6a46079c AK |
354 | |
355 | /* | |
3d7fed4a JC |
356 | * Send SIGKILL if "tk->addr == -EFAULT". Also, as |
357 | * "tk->size_shift" is always non-zero for !is_zone_device_page(), | |
358 | * so "tk->size_shift == 0" effectively checks no mapping on | |
359 | * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times | |
360 | * to a process' address space, it's possible not all N VMAs | |
361 | * contain mappings for the page, but at least one VMA does. | |
362 | * Only deliver SIGBUS with payload derived from the VMA that | |
363 | * has a mapping for the page. | |
6a46079c | 364 | */ |
3d7fed4a | 365 | if (tk->addr == -EFAULT) { |
495367c0 | 366 | pr_info("Memory failure: Unable to find user space address %lx in %s\n", |
6a46079c | 367 | page_to_pfn(p), tsk->comm); |
3d7fed4a JC |
368 | } else if (tk->size_shift == 0) { |
369 | kfree(tk); | |
370 | return; | |
6a46079c | 371 | } |
996ff7a0 | 372 | |
6a46079c AK |
373 | get_task_struct(tsk); |
374 | tk->tsk = tsk; | |
375 | list_add_tail(&tk->nd, to_kill); | |
376 | } | |
377 | ||
378 | /* | |
379 | * Kill the processes that have been collected earlier. | |
380 | * | |
381 | * Only do anything when DOIT is set, otherwise just free the list | |
382 | * (this is used for clean pages which do not need killing) | |
383 | * Also when FAIL is set do a force kill because something went | |
384 | * wrong earlier. | |
385 | */ | |
ae1139ec DW |
386 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
387 | unsigned long pfn, int flags) | |
6a46079c AK |
388 | { |
389 | struct to_kill *tk, *next; | |
390 | ||
391 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 392 | if (forcekill) { |
6a46079c | 393 | /* |
af901ca1 | 394 | * In case something went wrong with munmapping |
6a46079c AK |
395 | * make sure the process doesn't catch the |
396 | * signal and then access the memory. Just kill it. | |
6a46079c | 397 | */ |
3d7fed4a | 398 | if (fail || tk->addr == -EFAULT) { |
495367c0 | 399 | pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 400 | pfn, tk->tsk->comm, tk->tsk->pid); |
6376360e NH |
401 | do_send_sig_info(SIGKILL, SEND_SIG_PRIV, |
402 | tk->tsk, PIDTYPE_PID); | |
6a46079c AK |
403 | } |
404 | ||
405 | /* | |
406 | * In theory the process could have mapped | |
407 | * something else on the address in-between. We could | |
408 | * check for that, but we need to tell the | |
409 | * process anyways. | |
410 | */ | |
ae1139ec | 411 | else if (kill_proc(tk, pfn, flags) < 0) |
495367c0 | 412 | pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 413 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c AK |
414 | } |
415 | put_task_struct(tk->tsk); | |
416 | kfree(tk); | |
417 | } | |
418 | } | |
419 | ||
3ba08129 NH |
420 | /* |
421 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
422 | * on behalf of the thread group. Return task_struct of the (first found) | |
423 | * dedicated thread if found, and return NULL otherwise. | |
424 | * | |
425 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
426 | * have to call rcu_read_lock/unlock() in this function. | |
427 | */ | |
428 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 429 | { |
3ba08129 NH |
430 | struct task_struct *t; |
431 | ||
4e018b45 NH |
432 | for_each_thread(tsk, t) { |
433 | if (t->flags & PF_MCE_PROCESS) { | |
434 | if (t->flags & PF_MCE_EARLY) | |
435 | return t; | |
436 | } else { | |
437 | if (sysctl_memory_failure_early_kill) | |
438 | return t; | |
439 | } | |
440 | } | |
3ba08129 NH |
441 | return NULL; |
442 | } | |
443 | ||
444 | /* | |
445 | * Determine whether a given process is "early kill" process which expects | |
446 | * to be signaled when some page under the process is hwpoisoned. | |
447 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
30c9cf49 | 448 | * specified) if the process is "early kill" and otherwise returns NULL. |
03151c6e | 449 | * |
30c9cf49 AY |
450 | * Note that the above is true for Action Optional case. For Action Required |
451 | * case, it's only meaningful to the current thread which need to be signaled | |
452 | * with SIGBUS, this error is Action Optional for other non current | |
453 | * processes sharing the same error page,if the process is "early kill", the | |
454 | * task_struct of the dedicated thread will also be returned. | |
3ba08129 NH |
455 | */ |
456 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
457 | int force_early) | |
458 | { | |
6a46079c | 459 | if (!tsk->mm) |
3ba08129 | 460 | return NULL; |
30c9cf49 AY |
461 | /* |
462 | * Comparing ->mm here because current task might represent | |
463 | * a subthread, while tsk always points to the main thread. | |
464 | */ | |
465 | if (force_early && tsk->mm == current->mm) | |
466 | return current; | |
467 | ||
4e018b45 | 468 | return find_early_kill_thread(tsk); |
6a46079c AK |
469 | } |
470 | ||
471 | /* | |
472 | * Collect processes when the error hit an anonymous page. | |
473 | */ | |
474 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 475 | int force_early) |
6a46079c AK |
476 | { |
477 | struct vm_area_struct *vma; | |
478 | struct task_struct *tsk; | |
479 | struct anon_vma *av; | |
bf181b9f | 480 | pgoff_t pgoff; |
6a46079c | 481 | |
4fc3f1d6 | 482 | av = page_lock_anon_vma_read(page); |
6a46079c | 483 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
484 | return; |
485 | ||
a0f7a756 | 486 | pgoff = page_to_pgoff(page); |
9b679320 | 487 | read_lock(&tasklist_lock); |
6a46079c | 488 | for_each_process (tsk) { |
5beb4930 | 489 | struct anon_vma_chain *vmac; |
3ba08129 | 490 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 491 | |
3ba08129 | 492 | if (!t) |
6a46079c | 493 | continue; |
bf181b9f ML |
494 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
495 | pgoff, pgoff) { | |
5beb4930 | 496 | vma = vmac->vma; |
6a46079c AK |
497 | if (!page_mapped_in_vma(page, vma)) |
498 | continue; | |
3ba08129 | 499 | if (vma->vm_mm == t->mm) |
996ff7a0 | 500 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
501 | } |
502 | } | |
6a46079c | 503 | read_unlock(&tasklist_lock); |
4fc3f1d6 | 504 | page_unlock_anon_vma_read(av); |
6a46079c AK |
505 | } |
506 | ||
507 | /* | |
508 | * Collect processes when the error hit a file mapped page. | |
509 | */ | |
510 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 511 | int force_early) |
6a46079c AK |
512 | { |
513 | struct vm_area_struct *vma; | |
514 | struct task_struct *tsk; | |
6a46079c | 515 | struct address_space *mapping = page->mapping; |
c43bc03d | 516 | pgoff_t pgoff; |
6a46079c | 517 | |
d28eb9c8 | 518 | i_mmap_lock_read(mapping); |
9b679320 | 519 | read_lock(&tasklist_lock); |
c43bc03d | 520 | pgoff = page_to_pgoff(page); |
6a46079c | 521 | for_each_process(tsk) { |
3ba08129 | 522 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 523 | |
3ba08129 | 524 | if (!t) |
6a46079c | 525 | continue; |
6b2dbba8 | 526 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
527 | pgoff) { |
528 | /* | |
529 | * Send early kill signal to tasks where a vma covers | |
530 | * the page but the corrupted page is not necessarily | |
531 | * mapped it in its pte. | |
532 | * Assume applications who requested early kill want | |
533 | * to be informed of all such data corruptions. | |
534 | */ | |
3ba08129 | 535 | if (vma->vm_mm == t->mm) |
996ff7a0 | 536 | add_to_kill(t, page, vma, to_kill); |
6a46079c AK |
537 | } |
538 | } | |
6a46079c | 539 | read_unlock(&tasklist_lock); |
d28eb9c8 | 540 | i_mmap_unlock_read(mapping); |
6a46079c AK |
541 | } |
542 | ||
543 | /* | |
544 | * Collect the processes who have the corrupted page mapped to kill. | |
6a46079c | 545 | */ |
74614de1 TL |
546 | static void collect_procs(struct page *page, struct list_head *tokill, |
547 | int force_early) | |
6a46079c | 548 | { |
6a46079c AK |
549 | if (!page->mapping) |
550 | return; | |
551 | ||
6a46079c | 552 | if (PageAnon(page)) |
996ff7a0 | 553 | collect_procs_anon(page, tokill, force_early); |
6a46079c | 554 | else |
996ff7a0 | 555 | collect_procs_file(page, tokill, force_early); |
6a46079c AK |
556 | } |
557 | ||
a3f5d80e NH |
558 | struct hwp_walk { |
559 | struct to_kill tk; | |
560 | unsigned long pfn; | |
561 | int flags; | |
562 | }; | |
563 | ||
564 | static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift) | |
565 | { | |
566 | tk->addr = addr; | |
567 | tk->size_shift = shift; | |
568 | } | |
569 | ||
570 | static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift, | |
571 | unsigned long poisoned_pfn, struct to_kill *tk) | |
572 | { | |
573 | unsigned long pfn = 0; | |
574 | ||
575 | if (pte_present(pte)) { | |
576 | pfn = pte_pfn(pte); | |
577 | } else { | |
578 | swp_entry_t swp = pte_to_swp_entry(pte); | |
579 | ||
580 | if (is_hwpoison_entry(swp)) | |
581 | pfn = hwpoison_entry_to_pfn(swp); | |
582 | } | |
583 | ||
584 | if (!pfn || pfn != poisoned_pfn) | |
585 | return 0; | |
586 | ||
587 | set_to_kill(tk, addr, shift); | |
588 | return 1; | |
589 | } | |
590 | ||
591 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
592 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
593 | struct hwp_walk *hwp) | |
594 | { | |
595 | pmd_t pmd = *pmdp; | |
596 | unsigned long pfn; | |
597 | unsigned long hwpoison_vaddr; | |
598 | ||
599 | if (!pmd_present(pmd)) | |
600 | return 0; | |
601 | pfn = pmd_pfn(pmd); | |
602 | if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) { | |
603 | hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT); | |
604 | set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT); | |
605 | return 1; | |
606 | } | |
607 | return 0; | |
608 | } | |
609 | #else | |
610 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
611 | struct hwp_walk *hwp) | |
612 | { | |
613 | return 0; | |
614 | } | |
615 | #endif | |
616 | ||
617 | static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr, | |
618 | unsigned long end, struct mm_walk *walk) | |
619 | { | |
620 | struct hwp_walk *hwp = (struct hwp_walk *)walk->private; | |
621 | int ret = 0; | |
622 | pte_t *ptep; | |
623 | spinlock_t *ptl; | |
624 | ||
625 | ptl = pmd_trans_huge_lock(pmdp, walk->vma); | |
626 | if (ptl) { | |
627 | ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp); | |
628 | spin_unlock(ptl); | |
629 | goto out; | |
630 | } | |
631 | ||
632 | if (pmd_trans_unstable(pmdp)) | |
633 | goto out; | |
634 | ||
635 | ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, addr, &ptl); | |
636 | for (; addr != end; ptep++, addr += PAGE_SIZE) { | |
637 | ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT, | |
638 | hwp->pfn, &hwp->tk); | |
639 | if (ret == 1) | |
640 | break; | |
641 | } | |
642 | pte_unmap_unlock(ptep - 1, ptl); | |
643 | out: | |
644 | cond_resched(); | |
645 | return ret; | |
646 | } | |
647 | ||
648 | #ifdef CONFIG_HUGETLB_PAGE | |
649 | static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask, | |
650 | unsigned long addr, unsigned long end, | |
651 | struct mm_walk *walk) | |
652 | { | |
653 | struct hwp_walk *hwp = (struct hwp_walk *)walk->private; | |
654 | pte_t pte = huge_ptep_get(ptep); | |
655 | struct hstate *h = hstate_vma(walk->vma); | |
656 | ||
657 | return check_hwpoisoned_entry(pte, addr, huge_page_shift(h), | |
658 | hwp->pfn, &hwp->tk); | |
659 | } | |
660 | #else | |
661 | #define hwpoison_hugetlb_range NULL | |
662 | #endif | |
663 | ||
664 | static struct mm_walk_ops hwp_walk_ops = { | |
665 | .pmd_entry = hwpoison_pte_range, | |
666 | .hugetlb_entry = hwpoison_hugetlb_range, | |
667 | }; | |
668 | ||
669 | /* | |
670 | * Sends SIGBUS to the current process with error info. | |
671 | * | |
672 | * This function is intended to handle "Action Required" MCEs on already | |
673 | * hardware poisoned pages. They could happen, for example, when | |
674 | * memory_failure() failed to unmap the error page at the first call, or | |
675 | * when multiple local machine checks happened on different CPUs. | |
676 | * | |
677 | * MCE handler currently has no easy access to the error virtual address, | |
678 | * so this function walks page table to find it. The returned virtual address | |
679 | * is proper in most cases, but it could be wrong when the application | |
680 | * process has multiple entries mapping the error page. | |
681 | */ | |
682 | static int kill_accessing_process(struct task_struct *p, unsigned long pfn, | |
683 | int flags) | |
684 | { | |
685 | int ret; | |
686 | struct hwp_walk priv = { | |
687 | .pfn = pfn, | |
688 | }; | |
689 | priv.tk.tsk = p; | |
690 | ||
691 | mmap_read_lock(p->mm); | |
692 | ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops, | |
693 | (void *)&priv); | |
694 | if (ret == 1 && priv.tk.addr) | |
695 | kill_proc(&priv.tk, pfn, flags); | |
696 | mmap_read_unlock(p->mm); | |
697 | return ret ? -EFAULT : -EHWPOISON; | |
698 | } | |
699 | ||
6a46079c | 700 | static const char *action_name[] = { |
cc637b17 XX |
701 | [MF_IGNORED] = "Ignored", |
702 | [MF_FAILED] = "Failed", | |
703 | [MF_DELAYED] = "Delayed", | |
704 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
705 | }; |
706 | ||
707 | static const char * const action_page_types[] = { | |
cc637b17 XX |
708 | [MF_MSG_KERNEL] = "reserved kernel page", |
709 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
710 | [MF_MSG_SLAB] = "kernel slab page", | |
711 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
712 | [MF_MSG_POISONED_HUGE] = "huge page already hardware poisoned", | |
713 | [MF_MSG_HUGE] = "huge page", | |
714 | [MF_MSG_FREE_HUGE] = "free huge page", | |
31286a84 | 715 | [MF_MSG_NON_PMD_HUGE] = "non-pmd-sized huge page", |
cc637b17 XX |
716 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", |
717 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
718 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
719 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
720 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
721 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
722 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
723 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
724 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
725 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
726 | [MF_MSG_BUDDY] = "free buddy page", | |
727 | [MF_MSG_BUDDY_2ND] = "free buddy page (2nd try)", | |
6100e34b | 728 | [MF_MSG_DAX] = "dax page", |
5d1fd5dc | 729 | [MF_MSG_UNSPLIT_THP] = "unsplit thp", |
cc637b17 | 730 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
731 | }; |
732 | ||
dc2a1cbf WF |
733 | /* |
734 | * XXX: It is possible that a page is isolated from LRU cache, | |
735 | * and then kept in swap cache or failed to remove from page cache. | |
736 | * The page count will stop it from being freed by unpoison. | |
737 | * Stress tests should be aware of this memory leak problem. | |
738 | */ | |
739 | static int delete_from_lru_cache(struct page *p) | |
740 | { | |
741 | if (!isolate_lru_page(p)) { | |
742 | /* | |
743 | * Clear sensible page flags, so that the buddy system won't | |
744 | * complain when the page is unpoison-and-freed. | |
745 | */ | |
746 | ClearPageActive(p); | |
747 | ClearPageUnevictable(p); | |
18365225 MH |
748 | |
749 | /* | |
750 | * Poisoned page might never drop its ref count to 0 so we have | |
751 | * to uncharge it manually from its memcg. | |
752 | */ | |
753 | mem_cgroup_uncharge(p); | |
754 | ||
dc2a1cbf WF |
755 | /* |
756 | * drop the page count elevated by isolate_lru_page() | |
757 | */ | |
09cbfeaf | 758 | put_page(p); |
dc2a1cbf WF |
759 | return 0; |
760 | } | |
761 | return -EIO; | |
762 | } | |
763 | ||
78bb9203 NH |
764 | static int truncate_error_page(struct page *p, unsigned long pfn, |
765 | struct address_space *mapping) | |
766 | { | |
767 | int ret = MF_FAILED; | |
768 | ||
769 | if (mapping->a_ops->error_remove_page) { | |
770 | int err = mapping->a_ops->error_remove_page(mapping, p); | |
771 | ||
772 | if (err != 0) { | |
773 | pr_info("Memory failure: %#lx: Failed to punch page: %d\n", | |
774 | pfn, err); | |
775 | } else if (page_has_private(p) && | |
776 | !try_to_release_page(p, GFP_NOIO)) { | |
777 | pr_info("Memory failure: %#lx: failed to release buffers\n", | |
778 | pfn); | |
779 | } else { | |
780 | ret = MF_RECOVERED; | |
781 | } | |
782 | } else { | |
783 | /* | |
784 | * If the file system doesn't support it just invalidate | |
785 | * This fails on dirty or anything with private pages | |
786 | */ | |
787 | if (invalidate_inode_page(p)) | |
788 | ret = MF_RECOVERED; | |
789 | else | |
790 | pr_info("Memory failure: %#lx: Failed to invalidate\n", | |
791 | pfn); | |
792 | } | |
793 | ||
794 | return ret; | |
795 | } | |
796 | ||
6a46079c AK |
797 | /* |
798 | * Error hit kernel page. | |
799 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
800 | * could be more sophisticated. | |
801 | */ | |
802 | static int me_kernel(struct page *p, unsigned long pfn) | |
6a46079c | 803 | { |
ea6d0630 | 804 | unlock_page(p); |
cc637b17 | 805 | return MF_IGNORED; |
6a46079c AK |
806 | } |
807 | ||
808 | /* | |
809 | * Page in unknown state. Do nothing. | |
810 | */ | |
811 | static int me_unknown(struct page *p, unsigned long pfn) | |
812 | { | |
495367c0 | 813 | pr_err("Memory failure: %#lx: Unknown page state\n", pfn); |
ea6d0630 | 814 | unlock_page(p); |
cc637b17 | 815 | return MF_FAILED; |
6a46079c AK |
816 | } |
817 | ||
6a46079c AK |
818 | /* |
819 | * Clean (or cleaned) page cache page. | |
820 | */ | |
821 | static int me_pagecache_clean(struct page *p, unsigned long pfn) | |
822 | { | |
ea6d0630 | 823 | int ret; |
6a46079c AK |
824 | struct address_space *mapping; |
825 | ||
dc2a1cbf WF |
826 | delete_from_lru_cache(p); |
827 | ||
6a46079c AK |
828 | /* |
829 | * For anonymous pages we're done the only reference left | |
830 | * should be the one m_f() holds. | |
831 | */ | |
ea6d0630 NH |
832 | if (PageAnon(p)) { |
833 | ret = MF_RECOVERED; | |
834 | goto out; | |
835 | } | |
6a46079c AK |
836 | |
837 | /* | |
838 | * Now truncate the page in the page cache. This is really | |
839 | * more like a "temporary hole punch" | |
840 | * Don't do this for block devices when someone else | |
841 | * has a reference, because it could be file system metadata | |
842 | * and that's not safe to truncate. | |
843 | */ | |
844 | mapping = page_mapping(p); | |
845 | if (!mapping) { | |
846 | /* | |
847 | * Page has been teared down in the meanwhile | |
848 | */ | |
ea6d0630 NH |
849 | ret = MF_FAILED; |
850 | goto out; | |
6a46079c AK |
851 | } |
852 | ||
853 | /* | |
854 | * Truncation is a bit tricky. Enable it per file system for now. | |
855 | * | |
856 | * Open: to take i_mutex or not for this? Right now we don't. | |
857 | */ | |
ea6d0630 NH |
858 | ret = truncate_error_page(p, pfn, mapping); |
859 | out: | |
860 | unlock_page(p); | |
861 | return ret; | |
6a46079c AK |
862 | } |
863 | ||
864 | /* | |
549543df | 865 | * Dirty pagecache page |
6a46079c AK |
866 | * Issues: when the error hit a hole page the error is not properly |
867 | * propagated. | |
868 | */ | |
869 | static int me_pagecache_dirty(struct page *p, unsigned long pfn) | |
870 | { | |
871 | struct address_space *mapping = page_mapping(p); | |
872 | ||
873 | SetPageError(p); | |
874 | /* TBD: print more information about the file. */ | |
875 | if (mapping) { | |
876 | /* | |
877 | * IO error will be reported by write(), fsync(), etc. | |
878 | * who check the mapping. | |
879 | * This way the application knows that something went | |
880 | * wrong with its dirty file data. | |
881 | * | |
882 | * There's one open issue: | |
883 | * | |
884 | * The EIO will be only reported on the next IO | |
885 | * operation and then cleared through the IO map. | |
886 | * Normally Linux has two mechanisms to pass IO error | |
887 | * first through the AS_EIO flag in the address space | |
888 | * and then through the PageError flag in the page. | |
889 | * Since we drop pages on memory failure handling the | |
890 | * only mechanism open to use is through AS_AIO. | |
891 | * | |
892 | * This has the disadvantage that it gets cleared on | |
893 | * the first operation that returns an error, while | |
894 | * the PageError bit is more sticky and only cleared | |
895 | * when the page is reread or dropped. If an | |
896 | * application assumes it will always get error on | |
897 | * fsync, but does other operations on the fd before | |
25985edc | 898 | * and the page is dropped between then the error |
6a46079c AK |
899 | * will not be properly reported. |
900 | * | |
901 | * This can already happen even without hwpoisoned | |
902 | * pages: first on metadata IO errors (which only | |
903 | * report through AS_EIO) or when the page is dropped | |
904 | * at the wrong time. | |
905 | * | |
906 | * So right now we assume that the application DTRT on | |
907 | * the first EIO, but we're not worse than other parts | |
908 | * of the kernel. | |
909 | */ | |
af21bfaf | 910 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
911 | } |
912 | ||
913 | return me_pagecache_clean(p, pfn); | |
914 | } | |
915 | ||
916 | /* | |
917 | * Clean and dirty swap cache. | |
918 | * | |
919 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
920 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
921 | * referenced concurrently by 2 types of PTEs: | |
922 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
923 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
924 | * and then | |
925 | * - clear dirty bit to prevent IO | |
926 | * - remove from LRU | |
927 | * - but keep in the swap cache, so that when we return to it on | |
928 | * a later page fault, we know the application is accessing | |
929 | * corrupted data and shall be killed (we installed simple | |
930 | * interception code in do_swap_page to catch it). | |
931 | * | |
932 | * Clean swap cache pages can be directly isolated. A later page fault will | |
933 | * bring in the known good data from disk. | |
934 | */ | |
935 | static int me_swapcache_dirty(struct page *p, unsigned long pfn) | |
936 | { | |
ea6d0630 NH |
937 | int ret; |
938 | ||
6a46079c AK |
939 | ClearPageDirty(p); |
940 | /* Trigger EIO in shmem: */ | |
941 | ClearPageUptodate(p); | |
942 | ||
ea6d0630 NH |
943 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED; |
944 | unlock_page(p); | |
945 | return ret; | |
6a46079c AK |
946 | } |
947 | ||
948 | static int me_swapcache_clean(struct page *p, unsigned long pfn) | |
949 | { | |
ea6d0630 NH |
950 | int ret; |
951 | ||
6a46079c | 952 | delete_from_swap_cache(p); |
e43c3afb | 953 | |
ea6d0630 NH |
954 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED; |
955 | unlock_page(p); | |
956 | return ret; | |
6a46079c AK |
957 | } |
958 | ||
959 | /* | |
960 | * Huge pages. Needs work. | |
961 | * Issues: | |
93f70f90 NH |
962 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
963 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c AK |
964 | */ |
965 | static int me_huge_page(struct page *p, unsigned long pfn) | |
966 | { | |
a8b2c2ce | 967 | int res; |
93f70f90 | 968 | struct page *hpage = compound_head(p); |
78bb9203 | 969 | struct address_space *mapping; |
2491ffee NH |
970 | |
971 | if (!PageHuge(hpage)) | |
972 | return MF_DELAYED; | |
973 | ||
78bb9203 NH |
974 | mapping = page_mapping(hpage); |
975 | if (mapping) { | |
976 | res = truncate_error_page(hpage, pfn, mapping); | |
ea6d0630 | 977 | unlock_page(hpage); |
78bb9203 | 978 | } else { |
a8b2c2ce | 979 | res = MF_FAILED; |
78bb9203 NH |
980 | unlock_page(hpage); |
981 | /* | |
982 | * migration entry prevents later access on error anonymous | |
983 | * hugepage, so we can free and dissolve it into buddy to | |
984 | * save healthy subpages. | |
985 | */ | |
986 | if (PageAnon(hpage)) | |
987 | put_page(hpage); | |
a8b2c2ce OS |
988 | if (!dissolve_free_huge_page(p) && take_page_off_buddy(p)) { |
989 | page_ref_inc(p); | |
990 | res = MF_RECOVERED; | |
991 | } | |
93f70f90 | 992 | } |
78bb9203 NH |
993 | |
994 | return res; | |
6a46079c AK |
995 | } |
996 | ||
997 | /* | |
998 | * Various page states we can handle. | |
999 | * | |
1000 | * A page state is defined by its current page->flags bits. | |
1001 | * The table matches them in order and calls the right handler. | |
1002 | * | |
1003 | * This is quite tricky because we can access page at any time | |
25985edc | 1004 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
1005 | * |
1006 | * This is not complete. More states could be added. | |
1007 | * For any missing state don't attempt recovery. | |
1008 | */ | |
1009 | ||
1010 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 1011 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
1012 | #define unevict (1UL << PG_unevictable) |
1013 | #define mlock (1UL << PG_mlocked) | |
6a46079c | 1014 | #define lru (1UL << PG_lru) |
6a46079c | 1015 | #define head (1UL << PG_head) |
6a46079c | 1016 | #define slab (1UL << PG_slab) |
6a46079c AK |
1017 | #define reserved (1UL << PG_reserved) |
1018 | ||
1019 | static struct page_state { | |
1020 | unsigned long mask; | |
1021 | unsigned long res; | |
cc637b17 | 1022 | enum mf_action_page_type type; |
ea6d0630 NH |
1023 | |
1024 | /* Callback ->action() has to unlock the relevant page inside it. */ | |
6a46079c AK |
1025 | int (*action)(struct page *p, unsigned long pfn); |
1026 | } error_states[] = { | |
cc637b17 | 1027 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
1028 | /* |
1029 | * free pages are specially detected outside this table: | |
1030 | * PG_buddy pages only make a small fraction of all free pages. | |
1031 | */ | |
6a46079c AK |
1032 | |
1033 | /* | |
1034 | * Could in theory check if slab page is free or if we can drop | |
1035 | * currently unused objects without touching them. But just | |
1036 | * treat it as standard kernel for now. | |
1037 | */ | |
cc637b17 | 1038 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 1039 | |
cc637b17 | 1040 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 1041 | |
cc637b17 XX |
1042 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
1043 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 1044 | |
cc637b17 XX |
1045 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
1046 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 1047 | |
cc637b17 XX |
1048 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
1049 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 1050 | |
cc637b17 XX |
1051 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
1052 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
1053 | |
1054 | /* | |
1055 | * Catchall entry: must be at end. | |
1056 | */ | |
cc637b17 | 1057 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
1058 | }; |
1059 | ||
2326c467 AK |
1060 | #undef dirty |
1061 | #undef sc | |
1062 | #undef unevict | |
1063 | #undef mlock | |
2326c467 | 1064 | #undef lru |
2326c467 | 1065 | #undef head |
2326c467 AK |
1066 | #undef slab |
1067 | #undef reserved | |
1068 | ||
ff604cf6 NH |
1069 | /* |
1070 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
1071 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
1072 | */ | |
cc3e2af4 XX |
1073 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
1074 | enum mf_result result) | |
6a46079c | 1075 | { |
97f0b134 XX |
1076 | trace_memory_failure_event(pfn, type, result); |
1077 | ||
495367c0 | 1078 | pr_err("Memory failure: %#lx: recovery action for %s: %s\n", |
64d37a2b | 1079 | pfn, action_page_types[type], action_name[result]); |
6a46079c AK |
1080 | } |
1081 | ||
1082 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 1083 | unsigned long pfn) |
6a46079c AK |
1084 | { |
1085 | int result; | |
7456b040 | 1086 | int count; |
6a46079c | 1087 | |
ea6d0630 | 1088 | /* page p should be unlocked after returning from ps->action(). */ |
6a46079c | 1089 | result = ps->action(p, pfn); |
7456b040 | 1090 | |
bd1ce5f9 | 1091 | count = page_count(p) - 1; |
cc637b17 | 1092 | if (ps->action == me_swapcache_dirty && result == MF_DELAYED) |
138ce286 | 1093 | count--; |
78bb9203 | 1094 | if (count > 0) { |
495367c0 | 1095 | pr_err("Memory failure: %#lx: %s still referenced by %d users\n", |
64d37a2b | 1096 | pfn, action_page_types[ps->type], count); |
cc637b17 | 1097 | result = MF_FAILED; |
138ce286 | 1098 | } |
64d37a2b | 1099 | action_result(pfn, ps->type, result); |
6a46079c AK |
1100 | |
1101 | /* Could do more checks here if page looks ok */ | |
1102 | /* | |
1103 | * Could adjust zone counters here to correct for the missing page. | |
1104 | */ | |
1105 | ||
cc637b17 | 1106 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
1107 | } |
1108 | ||
25182f05 NH |
1109 | /* |
1110 | * Return true if a page type of a given page is supported by hwpoison | |
1111 | * mechanism (while handling could fail), otherwise false. This function | |
1112 | * does not return true for hugetlb or device memory pages, so it's assumed | |
1113 | * to be called only in the context where we never have such pages. | |
1114 | */ | |
1115 | static inline bool HWPoisonHandlable(struct page *page) | |
1116 | { | |
1117 | return PageLRU(page) || __PageMovable(page); | |
1118 | } | |
1119 | ||
ead07f6a | 1120 | /** |
17e395b6 | 1121 | * __get_hwpoison_page() - Get refcount for memory error handling: |
ead07f6a NH |
1122 | * @page: raw error page (hit by memory error) |
1123 | * | |
1124 | * Return: return 0 if failed to grab the refcount, otherwise true (some | |
1125 | * non-zero value.) | |
1126 | */ | |
17e395b6 | 1127 | static int __get_hwpoison_page(struct page *page) |
ead07f6a NH |
1128 | { |
1129 | struct page *head = compound_head(page); | |
25182f05 NH |
1130 | int ret = 0; |
1131 | bool hugetlb = false; | |
1132 | ||
1133 | ret = get_hwpoison_huge_page(head, &hugetlb); | |
1134 | if (hugetlb) | |
1135 | return ret; | |
1136 | ||
1137 | /* | |
1138 | * This check prevents from calling get_hwpoison_unless_zero() | |
1139 | * for any unsupported type of page in order to reduce the risk of | |
1140 | * unexpected races caused by taking a page refcount. | |
1141 | */ | |
1142 | if (!HWPoisonHandlable(head)) | |
1143 | return 0; | |
ead07f6a | 1144 | |
25182f05 | 1145 | if (PageTransHuge(head)) { |
98ed2b00 NH |
1146 | /* |
1147 | * Non anonymous thp exists only in allocation/free time. We | |
1148 | * can't handle such a case correctly, so let's give it up. | |
1149 | * This should be better than triggering BUG_ON when kernel | |
1150 | * tries to touch the "partially handled" page. | |
1151 | */ | |
1152 | if (!PageAnon(head)) { | |
495367c0 | 1153 | pr_err("Memory failure: %#lx: non anonymous thp\n", |
98ed2b00 NH |
1154 | page_to_pfn(page)); |
1155 | return 0; | |
1156 | } | |
ead07f6a NH |
1157 | } |
1158 | ||
c2e7e00b KK |
1159 | if (get_page_unless_zero(head)) { |
1160 | if (head == compound_head(page)) | |
1161 | return 1; | |
1162 | ||
495367c0 CY |
1163 | pr_info("Memory failure: %#lx cannot catch tail\n", |
1164 | page_to_pfn(page)); | |
c2e7e00b KK |
1165 | put_page(head); |
1166 | } | |
1167 | ||
1168 | return 0; | |
ead07f6a | 1169 | } |
ead07f6a | 1170 | |
2f714160 OS |
1171 | /* |
1172 | * Safely get reference count of an arbitrary page. | |
1173 | * | |
1174 | * Returns 0 for a free page, 1 for an in-use page, | |
1175 | * -EIO for a page-type we cannot handle and -EBUSY if we raced with an | |
1176 | * allocation. | |
1177 | * We only incremented refcount in case the page was already in-use and it | |
1178 | * is a known type we can handle. | |
1179 | */ | |
1180 | static int get_any_page(struct page *p, unsigned long flags) | |
17e395b6 | 1181 | { |
2f714160 OS |
1182 | int ret = 0, pass = 0; |
1183 | bool count_increased = false; | |
17e395b6 | 1184 | |
2f714160 OS |
1185 | if (flags & MF_COUNT_INCREASED) |
1186 | count_increased = true; | |
1187 | ||
1188 | try_again: | |
1189 | if (!count_increased && !__get_hwpoison_page(p)) { | |
1190 | if (page_count(p)) { | |
1191 | /* We raced with an allocation, retry. */ | |
1192 | if (pass++ < 3) | |
1193 | goto try_again; | |
1194 | ret = -EBUSY; | |
1195 | } else if (!PageHuge(p) && !is_free_buddy_page(p)) { | |
1196 | /* We raced with put_page, retry. */ | |
1197 | if (pass++ < 3) | |
1198 | goto try_again; | |
1199 | ret = -EIO; | |
1200 | } | |
1201 | } else { | |
25182f05 | 1202 | if (PageHuge(p) || HWPoisonHandlable(p)) { |
2f714160 OS |
1203 | ret = 1; |
1204 | } else { | |
1205 | /* | |
1206 | * A page we cannot handle. Check whether we can turn | |
1207 | * it into something we can handle. | |
1208 | */ | |
1209 | if (pass++ < 3) { | |
1210 | put_page(p); | |
1211 | shake_page(p, 1); | |
1212 | count_increased = false; | |
1213 | goto try_again; | |
1214 | } | |
1215 | put_page(p); | |
1216 | ret = -EIO; | |
1217 | } | |
17e395b6 OS |
1218 | } |
1219 | ||
1220 | return ret; | |
1221 | } | |
1222 | ||
2f714160 OS |
1223 | static int get_hwpoison_page(struct page *p, unsigned long flags, |
1224 | enum mf_flags ctxt) | |
1225 | { | |
1226 | int ret; | |
1227 | ||
1228 | zone_pcp_disable(page_zone(p)); | |
1229 | if (ctxt == MF_SOFT_OFFLINE) | |
1230 | ret = get_any_page(p, flags); | |
1231 | else | |
1232 | ret = __get_hwpoison_page(p); | |
1233 | zone_pcp_enable(page_zone(p)); | |
1234 | ||
1235 | return ret; | |
1236 | } | |
1237 | ||
6a46079c AK |
1238 | /* |
1239 | * Do all that is necessary to remove user space mappings. Unmap | |
1240 | * the pages and send SIGBUS to the processes if the data was dirty. | |
1241 | */ | |
666e5a40 | 1242 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
83b57531 | 1243 | int flags, struct page **hpagep) |
6a46079c | 1244 | { |
013339df | 1245 | enum ttu_flags ttu = TTU_IGNORE_MLOCK; |
6a46079c AK |
1246 | struct address_space *mapping; |
1247 | LIST_HEAD(tokill); | |
c0d0381a | 1248 | bool unmap_success = true; |
6751ed65 | 1249 | int kill = 1, forcekill; |
54b9dd14 | 1250 | struct page *hpage = *hpagep; |
286c469a | 1251 | bool mlocked = PageMlocked(hpage); |
6a46079c | 1252 | |
93a9eb39 NH |
1253 | /* |
1254 | * Here we are interested only in user-mapped pages, so skip any | |
1255 | * other types of pages. | |
1256 | */ | |
1257 | if (PageReserved(p) || PageSlab(p)) | |
666e5a40 | 1258 | return true; |
93a9eb39 | 1259 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 1260 | return true; |
6a46079c | 1261 | |
6a46079c AK |
1262 | /* |
1263 | * This check implies we don't kill processes if their pages | |
1264 | * are in the swap cache early. Those are always late kills. | |
1265 | */ | |
7af446a8 | 1266 | if (!page_mapped(hpage)) |
666e5a40 | 1267 | return true; |
1668bfd5 | 1268 | |
52089b14 | 1269 | if (PageKsm(p)) { |
495367c0 | 1270 | pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn); |
666e5a40 | 1271 | return false; |
52089b14 | 1272 | } |
6a46079c AK |
1273 | |
1274 | if (PageSwapCache(p)) { | |
495367c0 CY |
1275 | pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n", |
1276 | pfn); | |
6a46079c AK |
1277 | ttu |= TTU_IGNORE_HWPOISON; |
1278 | } | |
1279 | ||
1280 | /* | |
1281 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1282 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1283 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1284 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1285 | */ |
7af446a8 | 1286 | mapping = page_mapping(hpage); |
6751ed65 | 1287 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
f56753ac | 1288 | mapping_can_writeback(mapping)) { |
7af446a8 NH |
1289 | if (page_mkclean(hpage)) { |
1290 | SetPageDirty(hpage); | |
6a46079c AK |
1291 | } else { |
1292 | kill = 0; | |
1293 | ttu |= TTU_IGNORE_HWPOISON; | |
495367c0 | 1294 | pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1295 | pfn); |
1296 | } | |
1297 | } | |
1298 | ||
1299 | /* | |
1300 | * First collect all the processes that have the page | |
1301 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1302 | * because ttu takes the rmap data structures down. | |
1303 | * | |
1304 | * Error handling: We ignore errors here because | |
1305 | * there's nothing that can be done. | |
1306 | */ | |
1307 | if (kill) | |
415c64c1 | 1308 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1309 | |
c0d0381a MK |
1310 | if (!PageHuge(hpage)) { |
1311 | unmap_success = try_to_unmap(hpage, ttu); | |
1312 | } else { | |
336bf30e MK |
1313 | if (!PageAnon(hpage)) { |
1314 | /* | |
1315 | * For hugetlb pages in shared mappings, try_to_unmap | |
1316 | * could potentially call huge_pmd_unshare. Because of | |
1317 | * this, take semaphore in write mode here and set | |
1318 | * TTU_RMAP_LOCKED to indicate we have taken the lock | |
1319 | * at this higer level. | |
1320 | */ | |
1321 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1322 | if (mapping) { | |
1323 | unmap_success = try_to_unmap(hpage, | |
c0d0381a | 1324 | ttu|TTU_RMAP_LOCKED); |
336bf30e MK |
1325 | i_mmap_unlock_write(mapping); |
1326 | } else { | |
1327 | pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn); | |
1328 | unmap_success = false; | |
1329 | } | |
c0d0381a | 1330 | } else { |
336bf30e | 1331 | unmap_success = try_to_unmap(hpage, ttu); |
c0d0381a MK |
1332 | } |
1333 | } | |
666e5a40 | 1334 | if (!unmap_success) |
495367c0 | 1335 | pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1336 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1337 | |
286c469a NH |
1338 | /* |
1339 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1340 | * shake_page() again to ensure that it's flushed. | |
1341 | */ | |
1342 | if (mlocked) | |
1343 | shake_page(hpage, 0); | |
1344 | ||
6a46079c AK |
1345 | /* |
1346 | * Now that the dirty bit has been propagated to the | |
1347 | * struct page and all unmaps done we can decide if | |
1348 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1349 | * was dirty or the process is not restartable, |
1350 | * otherwise the tokill list is merely | |
6a46079c AK |
1351 | * freed. When there was a problem unmapping earlier |
1352 | * use a more force-full uncatchable kill to prevent | |
1353 | * any accesses to the poisoned memory. | |
1354 | */ | |
415c64c1 | 1355 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
ae1139ec | 1356 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1357 | |
666e5a40 | 1358 | return unmap_success; |
6a46079c AK |
1359 | } |
1360 | ||
0348d2eb NH |
1361 | static int identify_page_state(unsigned long pfn, struct page *p, |
1362 | unsigned long page_flags) | |
761ad8d7 NH |
1363 | { |
1364 | struct page_state *ps; | |
0348d2eb NH |
1365 | |
1366 | /* | |
1367 | * The first check uses the current page flags which may not have any | |
1368 | * relevant information. The second check with the saved page flags is | |
1369 | * carried out only if the first check can't determine the page status. | |
1370 | */ | |
1371 | for (ps = error_states;; ps++) | |
1372 | if ((p->flags & ps->mask) == ps->res) | |
1373 | break; | |
1374 | ||
1375 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1376 | ||
1377 | if (!ps->mask) | |
1378 | for (ps = error_states;; ps++) | |
1379 | if ((page_flags & ps->mask) == ps->res) | |
1380 | break; | |
1381 | return page_action(ps, p, pfn); | |
1382 | } | |
1383 | ||
694bf0b0 OS |
1384 | static int try_to_split_thp_page(struct page *page, const char *msg) |
1385 | { | |
1386 | lock_page(page); | |
1387 | if (!PageAnon(page) || unlikely(split_huge_page(page))) { | |
1388 | unsigned long pfn = page_to_pfn(page); | |
1389 | ||
1390 | unlock_page(page); | |
1391 | if (!PageAnon(page)) | |
1392 | pr_info("%s: %#lx: non anonymous thp\n", msg, pfn); | |
1393 | else | |
1394 | pr_info("%s: %#lx: thp split failed\n", msg, pfn); | |
1395 | put_page(page); | |
1396 | return -EBUSY; | |
1397 | } | |
1398 | unlock_page(page); | |
1399 | ||
1400 | return 0; | |
1401 | } | |
1402 | ||
83b57531 | 1403 | static int memory_failure_hugetlb(unsigned long pfn, int flags) |
0348d2eb | 1404 | { |
761ad8d7 NH |
1405 | struct page *p = pfn_to_page(pfn); |
1406 | struct page *head = compound_head(p); | |
1407 | int res; | |
1408 | unsigned long page_flags; | |
1409 | ||
1410 | if (TestSetPageHWPoison(head)) { | |
1411 | pr_err("Memory failure: %#lx: already hardware poisoned\n", | |
1412 | pfn); | |
a3f5d80e NH |
1413 | res = -EHWPOISON; |
1414 | if (flags & MF_ACTION_REQUIRED) | |
1415 | res = kill_accessing_process(current, page_to_pfn(head), flags); | |
1416 | return res; | |
761ad8d7 NH |
1417 | } |
1418 | ||
1419 | num_poisoned_pages_inc(); | |
1420 | ||
2f714160 | 1421 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p, flags, 0)) { |
761ad8d7 NH |
1422 | /* |
1423 | * Check "filter hit" and "race with other subpage." | |
1424 | */ | |
1425 | lock_page(head); | |
1426 | if (PageHWPoison(head)) { | |
1427 | if ((hwpoison_filter(p) && TestClearPageHWPoison(p)) | |
1428 | || (p != head && TestSetPageHWPoison(head))) { | |
1429 | num_poisoned_pages_dec(); | |
1430 | unlock_page(head); | |
1431 | return 0; | |
1432 | } | |
1433 | } | |
1434 | unlock_page(head); | |
a8b2c2ce OS |
1435 | res = MF_FAILED; |
1436 | if (!dissolve_free_huge_page(p) && take_page_off_buddy(p)) { | |
1437 | page_ref_inc(p); | |
1438 | res = MF_RECOVERED; | |
1439 | } | |
1440 | action_result(pfn, MF_MSG_FREE_HUGE, res); | |
1441 | return res == MF_RECOVERED ? 0 : -EBUSY; | |
761ad8d7 NH |
1442 | } |
1443 | ||
1444 | lock_page(head); | |
1445 | page_flags = head->flags; | |
1446 | ||
1447 | if (!PageHWPoison(head)) { | |
1448 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); | |
1449 | num_poisoned_pages_dec(); | |
1450 | unlock_page(head); | |
dd6e2402 | 1451 | put_page(head); |
761ad8d7 NH |
1452 | return 0; |
1453 | } | |
1454 | ||
31286a84 NH |
1455 | /* |
1456 | * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so | |
1457 | * simply disable it. In order to make it work properly, we need | |
1458 | * make sure that: | |
1459 | * - conversion of a pud that maps an error hugetlb into hwpoison | |
1460 | * entry properly works, and | |
1461 | * - other mm code walking over page table is aware of pud-aligned | |
1462 | * hwpoison entries. | |
1463 | */ | |
1464 | if (huge_page_size(page_hstate(head)) > PMD_SIZE) { | |
1465 | action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED); | |
1466 | res = -EBUSY; | |
1467 | goto out; | |
1468 | } | |
1469 | ||
83b57531 | 1470 | if (!hwpoison_user_mappings(p, pfn, flags, &head)) { |
761ad8d7 NH |
1471 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1472 | res = -EBUSY; | |
1473 | goto out; | |
1474 | } | |
1475 | ||
ea6d0630 | 1476 | return identify_page_state(pfn, p, page_flags); |
761ad8d7 NH |
1477 | out: |
1478 | unlock_page(head); | |
1479 | return res; | |
1480 | } | |
1481 | ||
6100e34b DW |
1482 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
1483 | struct dev_pagemap *pgmap) | |
1484 | { | |
1485 | struct page *page = pfn_to_page(pfn); | |
1486 | const bool unmap_success = true; | |
1487 | unsigned long size = 0; | |
1488 | struct to_kill *tk; | |
1489 | LIST_HEAD(tokill); | |
1490 | int rc = -EBUSY; | |
1491 | loff_t start; | |
27359fd6 | 1492 | dax_entry_t cookie; |
6100e34b | 1493 | |
1e8aaedb OS |
1494 | if (flags & MF_COUNT_INCREASED) |
1495 | /* | |
1496 | * Drop the extra refcount in case we come from madvise(). | |
1497 | */ | |
1498 | put_page(page); | |
1499 | ||
34dc45be DW |
1500 | /* device metadata space is not recoverable */ |
1501 | if (!pgmap_pfn_valid(pgmap, pfn)) { | |
1502 | rc = -ENXIO; | |
1503 | goto out; | |
1504 | } | |
1505 | ||
6100e34b DW |
1506 | /* |
1507 | * Prevent the inode from being freed while we are interrogating | |
1508 | * the address_space, typically this would be handled by | |
1509 | * lock_page(), but dax pages do not use the page lock. This | |
1510 | * also prevents changes to the mapping of this pfn until | |
1511 | * poison signaling is complete. | |
1512 | */ | |
27359fd6 MW |
1513 | cookie = dax_lock_page(page); |
1514 | if (!cookie) | |
6100e34b DW |
1515 | goto out; |
1516 | ||
1517 | if (hwpoison_filter(page)) { | |
1518 | rc = 0; | |
1519 | goto unlock; | |
1520 | } | |
1521 | ||
25b2995a | 1522 | if (pgmap->type == MEMORY_DEVICE_PRIVATE) { |
6100e34b DW |
1523 | /* |
1524 | * TODO: Handle HMM pages which may need coordination | |
1525 | * with device-side memory. | |
1526 | */ | |
1527 | goto unlock; | |
6100e34b DW |
1528 | } |
1529 | ||
1530 | /* | |
1531 | * Use this flag as an indication that the dax page has been | |
1532 | * remapped UC to prevent speculative consumption of poison. | |
1533 | */ | |
1534 | SetPageHWPoison(page); | |
1535 | ||
1536 | /* | |
1537 | * Unlike System-RAM there is no possibility to swap in a | |
1538 | * different physical page at a given virtual address, so all | |
1539 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1540 | * SIGBUS (i.e. MF_MUST_KILL) | |
1541 | */ | |
1542 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1543 | collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED); | |
1544 | ||
1545 | list_for_each_entry(tk, &tokill, nd) | |
1546 | if (tk->size_shift) | |
1547 | size = max(size, 1UL << tk->size_shift); | |
1548 | if (size) { | |
1549 | /* | |
1550 | * Unmap the largest mapping to avoid breaking up | |
1551 | * device-dax mappings which are constant size. The | |
1552 | * actual size of the mapping being torn down is | |
1553 | * communicated in siginfo, see kill_proc() | |
1554 | */ | |
1555 | start = (page->index << PAGE_SHIFT) & ~(size - 1); | |
4d75136b | 1556 | unmap_mapping_range(page->mapping, start, size, 0); |
6100e34b DW |
1557 | } |
1558 | kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags); | |
1559 | rc = 0; | |
1560 | unlock: | |
27359fd6 | 1561 | dax_unlock_page(page, cookie); |
6100e34b DW |
1562 | out: |
1563 | /* drop pgmap ref acquired in caller */ | |
1564 | put_dev_pagemap(pgmap); | |
1565 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
1566 | return rc; | |
1567 | } | |
1568 | ||
cd42f4a3 TL |
1569 | /** |
1570 | * memory_failure - Handle memory failure of a page. | |
1571 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
1572 | * @flags: fine tune action taken |
1573 | * | |
1574 | * This function is called by the low level machine check code | |
1575 | * of an architecture when it detects hardware memory corruption | |
1576 | * of a page. It tries its best to recover, which includes | |
1577 | * dropping pages, killing processes etc. | |
1578 | * | |
1579 | * The function is primarily of use for corruptions that | |
1580 | * happen outside the current execution context (e.g. when | |
1581 | * detected by a background scrubber) | |
1582 | * | |
1583 | * Must run in process context (e.g. a work queue) with interrupts | |
1584 | * enabled and no spinlocks hold. | |
1585 | */ | |
83b57531 | 1586 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 1587 | { |
6a46079c | 1588 | struct page *p; |
7af446a8 | 1589 | struct page *hpage; |
415c64c1 | 1590 | struct page *orig_head; |
6100e34b | 1591 | struct dev_pagemap *pgmap; |
171936dd | 1592 | int res = 0; |
524fca1e | 1593 | unsigned long page_flags; |
a8b2c2ce | 1594 | bool retry = true; |
171936dd | 1595 | static DEFINE_MUTEX(mf_mutex); |
6a46079c AK |
1596 | |
1597 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 1598 | panic("Memory failure on page %lx", pfn); |
6a46079c | 1599 | |
96c804a6 DH |
1600 | p = pfn_to_online_page(pfn); |
1601 | if (!p) { | |
1602 | if (pfn_valid(pfn)) { | |
1603 | pgmap = get_dev_pagemap(pfn, NULL); | |
1604 | if (pgmap) | |
1605 | return memory_failure_dev_pagemap(pfn, flags, | |
1606 | pgmap); | |
1607 | } | |
495367c0 CY |
1608 | pr_err("Memory failure: %#lx: memory outside kernel control\n", |
1609 | pfn); | |
a7560fc8 | 1610 | return -ENXIO; |
6a46079c AK |
1611 | } |
1612 | ||
171936dd TL |
1613 | mutex_lock(&mf_mutex); |
1614 | ||
a8b2c2ce | 1615 | try_again: |
171936dd TL |
1616 | if (PageHuge(p)) { |
1617 | res = memory_failure_hugetlb(pfn, flags); | |
1618 | goto unlock_mutex; | |
1619 | } | |
1620 | ||
6a46079c | 1621 | if (TestSetPageHWPoison(p)) { |
495367c0 CY |
1622 | pr_err("Memory failure: %#lx: already hardware poisoned\n", |
1623 | pfn); | |
47af12ba | 1624 | res = -EHWPOISON; |
a3f5d80e NH |
1625 | if (flags & MF_ACTION_REQUIRED) |
1626 | res = kill_accessing_process(current, pfn, flags); | |
171936dd | 1627 | goto unlock_mutex; |
6a46079c AK |
1628 | } |
1629 | ||
761ad8d7 | 1630 | orig_head = hpage = compound_head(p); |
b37ff71c | 1631 | num_poisoned_pages_inc(); |
6a46079c AK |
1632 | |
1633 | /* | |
1634 | * We need/can do nothing about count=0 pages. | |
1635 | * 1) it's a free page, and therefore in safe hand: | |
1636 | * prep_new_page() will be the gate keeper. | |
761ad8d7 | 1637 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
1638 | * Implies some kernel user: cannot stop them from |
1639 | * R/W the page; let's pray that the page has been | |
1640 | * used and will be freed some time later. | |
1641 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 1642 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 1643 | */ |
2f714160 | 1644 | if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p, flags, 0)) { |
8d22ba1b | 1645 | if (is_free_buddy_page(p)) { |
a8b2c2ce OS |
1646 | if (take_page_off_buddy(p)) { |
1647 | page_ref_inc(p); | |
1648 | res = MF_RECOVERED; | |
1649 | } else { | |
1650 | /* We lost the race, try again */ | |
1651 | if (retry) { | |
1652 | ClearPageHWPoison(p); | |
1653 | num_poisoned_pages_dec(); | |
1654 | retry = false; | |
1655 | goto try_again; | |
1656 | } | |
1657 | res = MF_FAILED; | |
1658 | } | |
1659 | action_result(pfn, MF_MSG_BUDDY, res); | |
171936dd | 1660 | res = res == MF_RECOVERED ? 0 : -EBUSY; |
8d22ba1b | 1661 | } else { |
cc637b17 | 1662 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); |
171936dd | 1663 | res = -EBUSY; |
8d22ba1b | 1664 | } |
171936dd | 1665 | goto unlock_mutex; |
6a46079c AK |
1666 | } |
1667 | ||
761ad8d7 | 1668 | if (PageTransHuge(hpage)) { |
5d1fd5dc NH |
1669 | if (try_to_split_thp_page(p, "Memory Failure") < 0) { |
1670 | action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED); | |
171936dd TL |
1671 | res = -EBUSY; |
1672 | goto unlock_mutex; | |
5d1fd5dc | 1673 | } |
415c64c1 | 1674 | VM_BUG_ON_PAGE(!page_count(p), p); |
415c64c1 NH |
1675 | } |
1676 | ||
e43c3afb WF |
1677 | /* |
1678 | * We ignore non-LRU pages for good reasons. | |
1679 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 1680 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
1681 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
1682 | * The check (unnecessarily) ignores LRU pages being isolated and | |
1683 | * walked by the page reclaim code, however that's not a big loss. | |
1684 | */ | |
8bcb74de | 1685 | shake_page(p, 0); |
e43c3afb | 1686 | |
761ad8d7 | 1687 | lock_page(p); |
847ce401 | 1688 | |
f37d4298 AK |
1689 | /* |
1690 | * The page could have changed compound pages during the locking. | |
1691 | * If this happens just bail out. | |
1692 | */ | |
415c64c1 | 1693 | if (PageCompound(p) && compound_head(p) != orig_head) { |
cc637b17 | 1694 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 | 1695 | res = -EBUSY; |
171936dd | 1696 | goto unlock_page; |
f37d4298 AK |
1697 | } |
1698 | ||
524fca1e NH |
1699 | /* |
1700 | * We use page flags to determine what action should be taken, but | |
1701 | * the flags can be modified by the error containment action. One | |
1702 | * example is an mlocked page, where PG_mlocked is cleared by | |
1703 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
1704 | * correctly, we save a copy of the page flags at this time. | |
1705 | */ | |
7d9d46ac | 1706 | page_flags = p->flags; |
524fca1e | 1707 | |
847ce401 WF |
1708 | /* |
1709 | * unpoison always clear PG_hwpoison inside page lock | |
1710 | */ | |
1711 | if (!PageHWPoison(p)) { | |
495367c0 | 1712 | pr_err("Memory failure: %#lx: just unpoisoned\n", pfn); |
b37ff71c | 1713 | num_poisoned_pages_dec(); |
761ad8d7 | 1714 | unlock_page(p); |
dd6e2402 | 1715 | put_page(p); |
171936dd | 1716 | goto unlock_mutex; |
847ce401 | 1717 | } |
7c116f2b WF |
1718 | if (hwpoison_filter(p)) { |
1719 | if (TestClearPageHWPoison(p)) | |
b37ff71c | 1720 | num_poisoned_pages_dec(); |
761ad8d7 | 1721 | unlock_page(p); |
dd6e2402 | 1722 | put_page(p); |
171936dd | 1723 | goto unlock_mutex; |
7c116f2b | 1724 | } |
847ce401 | 1725 | |
e8675d29 | 1726 | /* |
1727 | * __munlock_pagevec may clear a writeback page's LRU flag without | |
1728 | * page_lock. We need wait writeback completion for this page or it | |
1729 | * may trigger vfs BUG while evict inode. | |
1730 | */ | |
1731 | if (!PageTransTail(p) && !PageLRU(p) && !PageWriteback(p)) | |
0bc1f8b0 CY |
1732 | goto identify_page_state; |
1733 | ||
6edd6cc6 NH |
1734 | /* |
1735 | * It's very difficult to mess with pages currently under IO | |
1736 | * and in many cases impossible, so we just avoid it here. | |
1737 | */ | |
6a46079c AK |
1738 | wait_on_page_writeback(p); |
1739 | ||
1740 | /* | |
1741 | * Now take care of user space mappings. | |
e64a782f | 1742 | * Abort on fail: __delete_from_page_cache() assumes unmapped page. |
6a46079c | 1743 | */ |
1b473bec | 1744 | if (!hwpoison_user_mappings(p, pfn, flags, &p)) { |
cc637b17 | 1745 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 | 1746 | res = -EBUSY; |
171936dd | 1747 | goto unlock_page; |
1668bfd5 | 1748 | } |
6a46079c AK |
1749 | |
1750 | /* | |
1751 | * Torn down by someone else? | |
1752 | */ | |
dc2a1cbf | 1753 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 1754 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 1755 | res = -EBUSY; |
171936dd | 1756 | goto unlock_page; |
6a46079c AK |
1757 | } |
1758 | ||
0bc1f8b0 | 1759 | identify_page_state: |
0348d2eb | 1760 | res = identify_page_state(pfn, p, page_flags); |
ea6d0630 NH |
1761 | mutex_unlock(&mf_mutex); |
1762 | return res; | |
171936dd | 1763 | unlock_page: |
761ad8d7 | 1764 | unlock_page(p); |
171936dd TL |
1765 | unlock_mutex: |
1766 | mutex_unlock(&mf_mutex); | |
6a46079c AK |
1767 | return res; |
1768 | } | |
cd42f4a3 | 1769 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 1770 | |
ea8f5fb8 HY |
1771 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
1772 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
1773 | ||
1774 | struct memory_failure_entry { | |
1775 | unsigned long pfn; | |
ea8f5fb8 HY |
1776 | int flags; |
1777 | }; | |
1778 | ||
1779 | struct memory_failure_cpu { | |
1780 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
1781 | MEMORY_FAILURE_FIFO_SIZE); | |
1782 | spinlock_t lock; | |
1783 | struct work_struct work; | |
1784 | }; | |
1785 | ||
1786 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
1787 | ||
1788 | /** | |
1789 | * memory_failure_queue - Schedule handling memory failure of a page. | |
1790 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
1791 | * @flags: Flags for memory failure handling |
1792 | * | |
1793 | * This function is called by the low level hardware error handler | |
1794 | * when it detects hardware memory corruption of a page. It schedules | |
1795 | * the recovering of error page, including dropping pages, killing | |
1796 | * processes etc. | |
1797 | * | |
1798 | * The function is primarily of use for corruptions that | |
1799 | * happen outside the current execution context (e.g. when | |
1800 | * detected by a background scrubber) | |
1801 | * | |
1802 | * Can run in IRQ context. | |
1803 | */ | |
83b57531 | 1804 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
1805 | { |
1806 | struct memory_failure_cpu *mf_cpu; | |
1807 | unsigned long proc_flags; | |
1808 | struct memory_failure_entry entry = { | |
1809 | .pfn = pfn, | |
ea8f5fb8 HY |
1810 | .flags = flags, |
1811 | }; | |
1812 | ||
1813 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
1814 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 1815 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
1816 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
1817 | else | |
8e33a52f | 1818 | pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
1819 | pfn); |
1820 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1821 | put_cpu_var(memory_failure_cpu); | |
1822 | } | |
1823 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
1824 | ||
1825 | static void memory_failure_work_func(struct work_struct *work) | |
1826 | { | |
1827 | struct memory_failure_cpu *mf_cpu; | |
1828 | struct memory_failure_entry entry = { 0, }; | |
1829 | unsigned long proc_flags; | |
1830 | int gotten; | |
1831 | ||
06202231 | 1832 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
ea8f5fb8 HY |
1833 | for (;;) { |
1834 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
1835 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
1836 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
1837 | if (!gotten) | |
1838 | break; | |
cf870c70 | 1839 | if (entry.flags & MF_SOFT_OFFLINE) |
feec24a6 | 1840 | soft_offline_page(entry.pfn, entry.flags); |
cf870c70 | 1841 | else |
83b57531 | 1842 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
1843 | } |
1844 | } | |
1845 | ||
06202231 JM |
1846 | /* |
1847 | * Process memory_failure work queued on the specified CPU. | |
1848 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. | |
1849 | */ | |
1850 | void memory_failure_queue_kick(int cpu) | |
1851 | { | |
1852 | struct memory_failure_cpu *mf_cpu; | |
1853 | ||
1854 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1855 | cancel_work_sync(&mf_cpu->work); | |
1856 | memory_failure_work_func(&mf_cpu->work); | |
1857 | } | |
1858 | ||
ea8f5fb8 HY |
1859 | static int __init memory_failure_init(void) |
1860 | { | |
1861 | struct memory_failure_cpu *mf_cpu; | |
1862 | int cpu; | |
1863 | ||
1864 | for_each_possible_cpu(cpu) { | |
1865 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
1866 | spin_lock_init(&mf_cpu->lock); | |
1867 | INIT_KFIFO(mf_cpu->fifo); | |
1868 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
1869 | } | |
1870 | ||
1871 | return 0; | |
1872 | } | |
1873 | core_initcall(memory_failure_init); | |
1874 | ||
a5f65109 NH |
1875 | #define unpoison_pr_info(fmt, pfn, rs) \ |
1876 | ({ \ | |
1877 | if (__ratelimit(rs)) \ | |
1878 | pr_info(fmt, pfn); \ | |
1879 | }) | |
1880 | ||
847ce401 WF |
1881 | /** |
1882 | * unpoison_memory - Unpoison a previously poisoned page | |
1883 | * @pfn: Page number of the to be unpoisoned page | |
1884 | * | |
1885 | * Software-unpoison a page that has been poisoned by | |
1886 | * memory_failure() earlier. | |
1887 | * | |
1888 | * This is only done on the software-level, so it only works | |
1889 | * for linux injected failures, not real hardware failures | |
1890 | * | |
1891 | * Returns 0 for success, otherwise -errno. | |
1892 | */ | |
1893 | int unpoison_memory(unsigned long pfn) | |
1894 | { | |
1895 | struct page *page; | |
1896 | struct page *p; | |
1897 | int freeit = 0; | |
2f714160 | 1898 | unsigned long flags = 0; |
a5f65109 NH |
1899 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
1900 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
1901 | |
1902 | if (!pfn_valid(pfn)) | |
1903 | return -ENXIO; | |
1904 | ||
1905 | p = pfn_to_page(pfn); | |
1906 | page = compound_head(p); | |
1907 | ||
1908 | if (!PageHWPoison(p)) { | |
495367c0 | 1909 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 1910 | pfn, &unpoison_rs); |
847ce401 WF |
1911 | return 0; |
1912 | } | |
1913 | ||
230ac719 | 1914 | if (page_count(page) > 1) { |
495367c0 | 1915 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 1916 | pfn, &unpoison_rs); |
230ac719 NH |
1917 | return 0; |
1918 | } | |
1919 | ||
1920 | if (page_mapped(page)) { | |
495367c0 | 1921 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 1922 | pfn, &unpoison_rs); |
230ac719 NH |
1923 | return 0; |
1924 | } | |
1925 | ||
1926 | if (page_mapping(page)) { | |
495367c0 | 1927 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 1928 | pfn, &unpoison_rs); |
230ac719 NH |
1929 | return 0; |
1930 | } | |
1931 | ||
0cea3fdc WL |
1932 | /* |
1933 | * unpoison_memory() can encounter thp only when the thp is being | |
1934 | * worked by memory_failure() and the page lock is not held yet. | |
1935 | * In such case, we yield to memory_failure() and make unpoison fail. | |
1936 | */ | |
e76d30e2 | 1937 | if (!PageHuge(page) && PageTransHuge(page)) { |
495367c0 | 1938 | unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n", |
a5f65109 | 1939 | pfn, &unpoison_rs); |
ead07f6a | 1940 | return 0; |
0cea3fdc WL |
1941 | } |
1942 | ||
2f714160 | 1943 | if (!get_hwpoison_page(p, flags, 0)) { |
847ce401 | 1944 | if (TestClearPageHWPoison(p)) |
8e30456b | 1945 | num_poisoned_pages_dec(); |
495367c0 | 1946 | unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n", |
a5f65109 | 1947 | pfn, &unpoison_rs); |
847ce401 WF |
1948 | return 0; |
1949 | } | |
1950 | ||
7eaceacc | 1951 | lock_page(page); |
847ce401 WF |
1952 | /* |
1953 | * This test is racy because PG_hwpoison is set outside of page lock. | |
1954 | * That's acceptable because that won't trigger kernel panic. Instead, | |
1955 | * the PG_hwpoison page will be caught and isolated on the entrance to | |
1956 | * the free buddy page pool. | |
1957 | */ | |
c9fbdd5f | 1958 | if (TestClearPageHWPoison(page)) { |
495367c0 | 1959 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", |
a5f65109 | 1960 | pfn, &unpoison_rs); |
b37ff71c | 1961 | num_poisoned_pages_dec(); |
847ce401 WF |
1962 | freeit = 1; |
1963 | } | |
1964 | unlock_page(page); | |
1965 | ||
dd6e2402 | 1966 | put_page(page); |
3ba5eebc | 1967 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) |
dd6e2402 | 1968 | put_page(page); |
847ce401 WF |
1969 | |
1970 | return 0; | |
1971 | } | |
1972 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 1973 | |
6b9a217e | 1974 | static bool isolate_page(struct page *page, struct list_head *pagelist) |
d950b958 | 1975 | { |
6b9a217e OS |
1976 | bool isolated = false; |
1977 | bool lru = PageLRU(page); | |
d950b958 | 1978 | |
6b9a217e OS |
1979 | if (PageHuge(page)) { |
1980 | isolated = isolate_huge_page(page, pagelist); | |
1981 | } else { | |
1982 | if (lru) | |
1983 | isolated = !isolate_lru_page(page); | |
1984 | else | |
1985 | isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE); | |
1986 | ||
1987 | if (isolated) | |
1988 | list_add(&page->lru, pagelist); | |
0ebff32c | 1989 | } |
d950b958 | 1990 | |
6b9a217e OS |
1991 | if (isolated && lru) |
1992 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
1993 | page_is_file_lru(page)); | |
1994 | ||
03613808 | 1995 | /* |
6b9a217e OS |
1996 | * If we succeed to isolate the page, we grabbed another refcount on |
1997 | * the page, so we can safely drop the one we got from get_any_pages(). | |
1998 | * If we failed to isolate the page, it means that we cannot go further | |
1999 | * and we will return an error, so drop the reference we got from | |
2000 | * get_any_pages() as well. | |
03613808 | 2001 | */ |
6b9a217e OS |
2002 | put_page(page); |
2003 | return isolated; | |
d950b958 NH |
2004 | } |
2005 | ||
6b9a217e OS |
2006 | /* |
2007 | * __soft_offline_page handles hugetlb-pages and non-hugetlb pages. | |
2008 | * If the page is a non-dirty unmapped page-cache page, it simply invalidates. | |
2009 | * If the page is mapped, it migrates the contents over. | |
2010 | */ | |
2011 | static int __soft_offline_page(struct page *page) | |
af8fae7c | 2012 | { |
6b9a217e | 2013 | int ret = 0; |
af8fae7c | 2014 | unsigned long pfn = page_to_pfn(page); |
6b9a217e OS |
2015 | struct page *hpage = compound_head(page); |
2016 | char const *msg_page[] = {"page", "hugepage"}; | |
2017 | bool huge = PageHuge(page); | |
2018 | LIST_HEAD(pagelist); | |
54608759 JK |
2019 | struct migration_target_control mtc = { |
2020 | .nid = NUMA_NO_NODE, | |
2021 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
2022 | }; | |
facb6011 | 2023 | |
facb6011 | 2024 | /* |
af8fae7c NH |
2025 | * Check PageHWPoison again inside page lock because PageHWPoison |
2026 | * is set by memory_failure() outside page lock. Note that | |
2027 | * memory_failure() also double-checks PageHWPoison inside page lock, | |
2028 | * so there's no race between soft_offline_page() and memory_failure(). | |
facb6011 | 2029 | */ |
0ebff32c | 2030 | lock_page(page); |
6b9a217e OS |
2031 | if (!PageHuge(page)) |
2032 | wait_on_page_writeback(page); | |
af8fae7c NH |
2033 | if (PageHWPoison(page)) { |
2034 | unlock_page(page); | |
dd6e2402 | 2035 | put_page(page); |
af8fae7c | 2036 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
5a2ffca3 | 2037 | return 0; |
af8fae7c | 2038 | } |
6b9a217e OS |
2039 | |
2040 | if (!PageHuge(page)) | |
2041 | /* | |
2042 | * Try to invalidate first. This should work for | |
2043 | * non dirty unmapped page cache pages. | |
2044 | */ | |
2045 | ret = invalidate_inode_page(page); | |
facb6011 | 2046 | unlock_page(page); |
6b9a217e | 2047 | |
facb6011 | 2048 | /* |
facb6011 AK |
2049 | * RED-PEN would be better to keep it isolated here, but we |
2050 | * would need to fix isolation locking first. | |
2051 | */ | |
6b9a217e | 2052 | if (ret) { |
fb46e735 | 2053 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
6b9a217e | 2054 | page_handle_poison(page, false, true); |
af8fae7c | 2055 | return 0; |
facb6011 AK |
2056 | } |
2057 | ||
6b9a217e | 2058 | if (isolate_page(hpage, &pagelist)) { |
54608759 JK |
2059 | ret = migrate_pages(&pagelist, alloc_migration_target, NULL, |
2060 | (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE); | |
79f5f8fa | 2061 | if (!ret) { |
6b9a217e OS |
2062 | bool release = !huge; |
2063 | ||
2064 | if (!page_handle_poison(page, huge, release)) | |
2065 | ret = -EBUSY; | |
79f5f8fa | 2066 | } else { |
85fbe5d1 YX |
2067 | if (!list_empty(&pagelist)) |
2068 | putback_movable_pages(&pagelist); | |
59c82b70 | 2069 | |
6b9a217e OS |
2070 | pr_info("soft offline: %#lx: %s migration failed %d, type %lx (%pGp)\n", |
2071 | pfn, msg_page[huge], ret, page->flags, &page->flags); | |
facb6011 | 2072 | if (ret > 0) |
3f4b815a | 2073 | ret = -EBUSY; |
facb6011 AK |
2074 | } |
2075 | } else { | |
3f4b815a OS |
2076 | pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %lx (%pGp)\n", |
2077 | pfn, msg_page[huge], page_count(page), page->flags, &page->flags); | |
6b9a217e | 2078 | ret = -EBUSY; |
facb6011 | 2079 | } |
facb6011 AK |
2080 | return ret; |
2081 | } | |
86e05773 | 2082 | |
6b9a217e | 2083 | static int soft_offline_in_use_page(struct page *page) |
acc14dc4 | 2084 | { |
acc14dc4 NH |
2085 | struct page *hpage = compound_head(page); |
2086 | ||
694bf0b0 OS |
2087 | if (!PageHuge(page) && PageTransHuge(hpage)) |
2088 | if (try_to_split_thp_page(page, "soft offline") < 0) | |
acc14dc4 | 2089 | return -EBUSY; |
6b9a217e | 2090 | return __soft_offline_page(page); |
acc14dc4 NH |
2091 | } |
2092 | ||
d4ae9916 | 2093 | static int soft_offline_free_page(struct page *page) |
acc14dc4 | 2094 | { |
6b9a217e | 2095 | int rc = 0; |
acc14dc4 | 2096 | |
6b9a217e OS |
2097 | if (!page_handle_poison(page, true, false)) |
2098 | rc = -EBUSY; | |
06be6ff3 | 2099 | |
d4ae9916 | 2100 | return rc; |
acc14dc4 NH |
2101 | } |
2102 | ||
dad4e5b3 DW |
2103 | static void put_ref_page(struct page *page) |
2104 | { | |
2105 | if (page) | |
2106 | put_page(page); | |
2107 | } | |
2108 | ||
86e05773 WL |
2109 | /** |
2110 | * soft_offline_page - Soft offline a page. | |
feec24a6 | 2111 | * @pfn: pfn to soft-offline |
86e05773 WL |
2112 | * @flags: flags. Same as memory_failure(). |
2113 | * | |
2114 | * Returns 0 on success, otherwise negated errno. | |
2115 | * | |
2116 | * Soft offline a page, by migration or invalidation, | |
2117 | * without killing anything. This is for the case when | |
2118 | * a page is not corrupted yet (so it's still valid to access), | |
2119 | * but has had a number of corrected errors and is better taken | |
2120 | * out. | |
2121 | * | |
2122 | * The actual policy on when to do that is maintained by | |
2123 | * user space. | |
2124 | * | |
2125 | * This should never impact any application or cause data loss, | |
2126 | * however it might take some time. | |
2127 | * | |
2128 | * This is not a 100% solution for all memory, but tries to be | |
2129 | * ``good enough'' for the majority of memory. | |
2130 | */ | |
feec24a6 | 2131 | int soft_offline_page(unsigned long pfn, int flags) |
86e05773 WL |
2132 | { |
2133 | int ret; | |
b94e0282 | 2134 | bool try_again = true; |
dad4e5b3 DW |
2135 | struct page *page, *ref_page = NULL; |
2136 | ||
2137 | WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED)); | |
86e05773 | 2138 | |
feec24a6 NH |
2139 | if (!pfn_valid(pfn)) |
2140 | return -ENXIO; | |
dad4e5b3 DW |
2141 | if (flags & MF_COUNT_INCREASED) |
2142 | ref_page = pfn_to_page(pfn); | |
2143 | ||
feec24a6 NH |
2144 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ |
2145 | page = pfn_to_online_page(pfn); | |
dad4e5b3 DW |
2146 | if (!page) { |
2147 | put_ref_page(ref_page); | |
86a66810 | 2148 | return -EIO; |
dad4e5b3 | 2149 | } |
86a66810 | 2150 | |
86e05773 | 2151 | if (PageHWPoison(page)) { |
8295d535 | 2152 | pr_info("%s: %#lx page already poisoned\n", __func__, pfn); |
dad4e5b3 | 2153 | put_ref_page(ref_page); |
5a2ffca3 | 2154 | return 0; |
86e05773 | 2155 | } |
86e05773 | 2156 | |
b94e0282 | 2157 | retry: |
bfc8c901 | 2158 | get_online_mems(); |
2f714160 | 2159 | ret = get_hwpoison_page(page, flags, MF_SOFT_OFFLINE); |
bfc8c901 | 2160 | put_online_mems(); |
4e41a30c | 2161 | |
8295d535 | 2162 | if (ret > 0) { |
6b9a217e | 2163 | ret = soft_offline_in_use_page(page); |
8295d535 | 2164 | } else if (ret == 0) { |
b94e0282 OS |
2165 | if (soft_offline_free_page(page) && try_again) { |
2166 | try_again = false; | |
2167 | goto retry; | |
2168 | } | |
8295d535 | 2169 | } else if (ret == -EIO) { |
6696d2a6 | 2170 | pr_info("%s: %#lx: unknown page type: %lx (%pGp)\n", |
8295d535 OS |
2171 | __func__, pfn, page->flags, &page->flags); |
2172 | } | |
4e41a30c | 2173 | |
86e05773 WL |
2174 | return ret; |
2175 | } |