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 | */ |
96f96763 KW |
36 | |
37 | #define pr_fmt(fmt) "Memory failure: " fmt | |
38 | ||
6a46079c AK |
39 | #include <linux/kernel.h> |
40 | #include <linux/mm.h> | |
41 | #include <linux/page-flags.h> | |
478c5ffc | 42 | #include <linux/kernel-page-flags.h> |
3f07c014 | 43 | #include <linux/sched/signal.h> |
29930025 | 44 | #include <linux/sched/task.h> |
96c84dde | 45 | #include <linux/dax.h> |
01e00f88 | 46 | #include <linux/ksm.h> |
6a46079c | 47 | #include <linux/rmap.h> |
b9e15baf | 48 | #include <linux/export.h> |
6a46079c AK |
49 | #include <linux/pagemap.h> |
50 | #include <linux/swap.h> | |
51 | #include <linux/backing-dev.h> | |
facb6011 | 52 | #include <linux/migrate.h> |
facb6011 | 53 | #include <linux/suspend.h> |
5a0e3ad6 | 54 | #include <linux/slab.h> |
bf998156 | 55 | #include <linux/swapops.h> |
7af446a8 | 56 | #include <linux/hugetlb.h> |
20d6c96b | 57 | #include <linux/memory_hotplug.h> |
5db8a73a | 58 | #include <linux/mm_inline.h> |
6100e34b | 59 | #include <linux/memremap.h> |
ea8f5fb8 | 60 | #include <linux/kfifo.h> |
a5f65109 | 61 | #include <linux/ratelimit.h> |
d4ae9916 | 62 | #include <linux/page-isolation.h> |
a3f5d80e | 63 | #include <linux/pagewalk.h> |
a7605426 | 64 | #include <linux/shmem_fs.h> |
014bb1de | 65 | #include "swap.h" |
6a46079c | 66 | #include "internal.h" |
97f0b134 | 67 | #include "ras/ras_event.h" |
6a46079c AK |
68 | |
69 | int sysctl_memory_failure_early_kill __read_mostly = 0; | |
70 | ||
71 | int sysctl_memory_failure_recovery __read_mostly = 1; | |
72 | ||
293c07e3 | 73 | atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0); |
6a46079c | 74 | |
67f22ba7 | 75 | static bool hw_memory_failure __read_mostly = false; |
76 | ||
7453bf62 NH |
77 | /* |
78 | * Return values: | |
79 | * 1: the page is dissolved (if needed) and taken off from buddy, | |
80 | * 0: the page is dissolved (if needed) and not taken off from buddy, | |
81 | * < 0: failed to dissolve. | |
82 | */ | |
83 | static int __page_handle_poison(struct page *page) | |
510d25c9 | 84 | { |
f87060d3 | 85 | int ret; |
510d25c9 NH |
86 | |
87 | zone_pcp_disable(page_zone(page)); | |
88 | ret = dissolve_free_huge_page(page); | |
89 | if (!ret) | |
90 | ret = take_page_off_buddy(page); | |
91 | zone_pcp_enable(page_zone(page)); | |
92 | ||
7453bf62 | 93 | return ret; |
510d25c9 NH |
94 | } |
95 | ||
6b9a217e | 96 | static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release) |
06be6ff3 | 97 | { |
6b9a217e OS |
98 | if (hugepage_or_freepage) { |
99 | /* | |
100 | * Doing this check for free pages is also fine since dissolve_free_huge_page | |
101 | * returns 0 for non-hugetlb pages as well. | |
102 | */ | |
7453bf62 | 103 | if (__page_handle_poison(page) <= 0) |
6b9a217e OS |
104 | /* |
105 | * We could fail to take off the target page from buddy | |
f0953a1b | 106 | * for example due to racy page allocation, but that's |
6b9a217e OS |
107 | * acceptable because soft-offlined page is not broken |
108 | * and if someone really want to use it, they should | |
109 | * take it. | |
110 | */ | |
111 | return false; | |
112 | } | |
113 | ||
06be6ff3 | 114 | SetPageHWPoison(page); |
79f5f8fa OS |
115 | if (release) |
116 | put_page(page); | |
06be6ff3 OS |
117 | page_ref_inc(page); |
118 | num_poisoned_pages_inc(); | |
6b9a217e OS |
119 | |
120 | return true; | |
06be6ff3 OS |
121 | } |
122 | ||
27df5068 AK |
123 | #if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE) |
124 | ||
1bfe5feb | 125 | u32 hwpoison_filter_enable = 0; |
7c116f2b WF |
126 | u32 hwpoison_filter_dev_major = ~0U; |
127 | u32 hwpoison_filter_dev_minor = ~0U; | |
478c5ffc WF |
128 | u64 hwpoison_filter_flags_mask; |
129 | u64 hwpoison_filter_flags_value; | |
1bfe5feb | 130 | EXPORT_SYMBOL_GPL(hwpoison_filter_enable); |
7c116f2b WF |
131 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major); |
132 | EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor); | |
478c5ffc WF |
133 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask); |
134 | EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value); | |
7c116f2b WF |
135 | |
136 | static int hwpoison_filter_dev(struct page *p) | |
137 | { | |
138 | struct address_space *mapping; | |
139 | dev_t dev; | |
140 | ||
141 | if (hwpoison_filter_dev_major == ~0U && | |
142 | hwpoison_filter_dev_minor == ~0U) | |
143 | return 0; | |
144 | ||
7c116f2b WF |
145 | mapping = page_mapping(p); |
146 | if (mapping == NULL || mapping->host == NULL) | |
147 | return -EINVAL; | |
148 | ||
149 | dev = mapping->host->i_sb->s_dev; | |
150 | if (hwpoison_filter_dev_major != ~0U && | |
151 | hwpoison_filter_dev_major != MAJOR(dev)) | |
152 | return -EINVAL; | |
153 | if (hwpoison_filter_dev_minor != ~0U && | |
154 | hwpoison_filter_dev_minor != MINOR(dev)) | |
155 | return -EINVAL; | |
156 | ||
157 | return 0; | |
158 | } | |
159 | ||
478c5ffc WF |
160 | static int hwpoison_filter_flags(struct page *p) |
161 | { | |
162 | if (!hwpoison_filter_flags_mask) | |
163 | return 0; | |
164 | ||
165 | if ((stable_page_flags(p) & hwpoison_filter_flags_mask) == | |
166 | hwpoison_filter_flags_value) | |
167 | return 0; | |
168 | else | |
169 | return -EINVAL; | |
170 | } | |
171 | ||
4fd466eb AK |
172 | /* |
173 | * This allows stress tests to limit test scope to a collection of tasks | |
174 | * by putting them under some memcg. This prevents killing unrelated/important | |
175 | * processes such as /sbin/init. Note that the target task may share clean | |
176 | * pages with init (eg. libc text), which is harmless. If the target task | |
177 | * share _dirty_ pages with another task B, the test scheme must make sure B | |
178 | * is also included in the memcg. At last, due to race conditions this filter | |
179 | * can only guarantee that the page either belongs to the memcg tasks, or is | |
180 | * a freed page. | |
181 | */ | |
94a59fb3 | 182 | #ifdef CONFIG_MEMCG |
4fd466eb AK |
183 | u64 hwpoison_filter_memcg; |
184 | EXPORT_SYMBOL_GPL(hwpoison_filter_memcg); | |
185 | static int hwpoison_filter_task(struct page *p) | |
186 | { | |
4fd466eb AK |
187 | if (!hwpoison_filter_memcg) |
188 | return 0; | |
189 | ||
94a59fb3 | 190 | if (page_cgroup_ino(p) != hwpoison_filter_memcg) |
4fd466eb AK |
191 | return -EINVAL; |
192 | ||
193 | return 0; | |
194 | } | |
195 | #else | |
196 | static int hwpoison_filter_task(struct page *p) { return 0; } | |
197 | #endif | |
198 | ||
7c116f2b WF |
199 | int hwpoison_filter(struct page *p) |
200 | { | |
1bfe5feb HL |
201 | if (!hwpoison_filter_enable) |
202 | return 0; | |
203 | ||
7c116f2b WF |
204 | if (hwpoison_filter_dev(p)) |
205 | return -EINVAL; | |
206 | ||
478c5ffc WF |
207 | if (hwpoison_filter_flags(p)) |
208 | return -EINVAL; | |
209 | ||
4fd466eb AK |
210 | if (hwpoison_filter_task(p)) |
211 | return -EINVAL; | |
212 | ||
7c116f2b WF |
213 | return 0; |
214 | } | |
27df5068 AK |
215 | #else |
216 | int hwpoison_filter(struct page *p) | |
217 | { | |
218 | return 0; | |
219 | } | |
220 | #endif | |
221 | ||
7c116f2b WF |
222 | EXPORT_SYMBOL_GPL(hwpoison_filter); |
223 | ||
ae1139ec DW |
224 | /* |
225 | * Kill all processes that have a poisoned page mapped and then isolate | |
226 | * the page. | |
227 | * | |
228 | * General strategy: | |
229 | * Find all processes having the page mapped and kill them. | |
230 | * But we keep a page reference around so that the page is not | |
231 | * actually freed yet. | |
232 | * Then stash the page away | |
233 | * | |
234 | * There's no convenient way to get back to mapped processes | |
235 | * from the VMAs. So do a brute-force search over all | |
236 | * running processes. | |
237 | * | |
238 | * Remember that machine checks are not common (or rather | |
239 | * if they are common you have other problems), so this shouldn't | |
240 | * be a performance issue. | |
241 | * | |
242 | * Also there are some races possible while we get from the | |
243 | * error detection to actually handle it. | |
244 | */ | |
245 | ||
246 | struct to_kill { | |
247 | struct list_head nd; | |
248 | struct task_struct *tsk; | |
249 | unsigned long addr; | |
250 | short size_shift; | |
ae1139ec DW |
251 | }; |
252 | ||
6a46079c | 253 | /* |
7329bbeb TL |
254 | * Send all the processes who have the page mapped a signal. |
255 | * ``action optional'' if they are not immediately affected by the error | |
256 | * ``action required'' if error happened in current execution context | |
6a46079c | 257 | */ |
ae1139ec | 258 | static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags) |
6a46079c | 259 | { |
ae1139ec DW |
260 | struct task_struct *t = tk->tsk; |
261 | short addr_lsb = tk->size_shift; | |
872e9a20 | 262 | int ret = 0; |
6a46079c | 263 | |
96f96763 | 264 | pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n", |
872e9a20 | 265 | pfn, t->comm, t->pid); |
7329bbeb | 266 | |
49775047 ML |
267 | if ((flags & MF_ACTION_REQUIRED) && (t == current)) |
268 | ret = force_sig_mceerr(BUS_MCEERR_AR, | |
269 | (void __user *)tk->addr, addr_lsb); | |
270 | else | |
7329bbeb | 271 | /* |
49775047 ML |
272 | * Signal other processes sharing the page if they have |
273 | * PF_MCE_EARLY set. | |
7329bbeb TL |
274 | * Don't use force here, it's convenient if the signal |
275 | * can be temporarily blocked. | |
276 | * This could cause a loop when the user sets SIGBUS | |
277 | * to SIG_IGN, but hopefully no one will do that? | |
278 | */ | |
ae1139ec | 279 | ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr, |
c0f45555 | 280 | addr_lsb, t); /* synchronous? */ |
6a46079c | 281 | if (ret < 0) |
96f96763 | 282 | pr_info("Error sending signal to %s:%d: %d\n", |
1170532b | 283 | t->comm, t->pid, ret); |
6a46079c AK |
284 | return ret; |
285 | } | |
286 | ||
588f9ce6 | 287 | /* |
47e431f4 | 288 | * Unknown page type encountered. Try to check whether it can turn PageLRU by |
d0505e9f | 289 | * lru_add_drain_all. |
588f9ce6 | 290 | */ |
d0505e9f | 291 | void shake_page(struct page *p) |
588f9ce6 | 292 | { |
8bcb74de NH |
293 | if (PageHuge(p)) |
294 | return; | |
295 | ||
588f9ce6 AK |
296 | if (!PageSlab(p)) { |
297 | lru_add_drain_all(); | |
588f9ce6 AK |
298 | if (PageLRU(p) || is_free_buddy_page(p)) |
299 | return; | |
300 | } | |
facb6011 | 301 | |
588f9ce6 | 302 | /* |
d0505e9f YS |
303 | * TODO: Could shrink slab caches here if a lightweight range-based |
304 | * shrinker will be available. | |
588f9ce6 AK |
305 | */ |
306 | } | |
307 | EXPORT_SYMBOL_GPL(shake_page); | |
308 | ||
c36e2024 SR |
309 | static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct *vma, |
310 | unsigned long address) | |
6100e34b | 311 | { |
5c91c0e7 | 312 | unsigned long ret = 0; |
6100e34b DW |
313 | pgd_t *pgd; |
314 | p4d_t *p4d; | |
315 | pud_t *pud; | |
316 | pmd_t *pmd; | |
317 | pte_t *pte; | |
318 | ||
a994402b | 319 | VM_BUG_ON_VMA(address == -EFAULT, vma); |
6100e34b DW |
320 | pgd = pgd_offset(vma->vm_mm, address); |
321 | if (!pgd_present(*pgd)) | |
322 | return 0; | |
323 | p4d = p4d_offset(pgd, address); | |
324 | if (!p4d_present(*p4d)) | |
325 | return 0; | |
326 | pud = pud_offset(p4d, address); | |
327 | if (!pud_present(*pud)) | |
328 | return 0; | |
329 | if (pud_devmap(*pud)) | |
330 | return PUD_SHIFT; | |
331 | pmd = pmd_offset(pud, address); | |
332 | if (!pmd_present(*pmd)) | |
333 | return 0; | |
334 | if (pmd_devmap(*pmd)) | |
335 | return PMD_SHIFT; | |
336 | pte = pte_offset_map(pmd, address); | |
5c91c0e7 QZ |
337 | if (pte_present(*pte) && pte_devmap(*pte)) |
338 | ret = PAGE_SHIFT; | |
339 | pte_unmap(pte); | |
340 | return ret; | |
6100e34b | 341 | } |
6a46079c AK |
342 | |
343 | /* | |
344 | * Failure handling: if we can't find or can't kill a process there's | |
345 | * not much we can do. We just print a message and ignore otherwise. | |
346 | */ | |
347 | ||
ac87ca0e DW |
348 | #define FSDAX_INVALID_PGOFF ULONG_MAX |
349 | ||
6a46079c AK |
350 | /* |
351 | * Schedule a process for later kill. | |
352 | * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM. | |
c36e2024 | 353 | * |
ac87ca0e DW |
354 | * Note: @fsdax_pgoff is used only when @p is a fsdax page and a |
355 | * filesystem with a memory failure handler has claimed the | |
356 | * memory_failure event. In all other cases, page->index and | |
357 | * page->mapping are sufficient for mapping the page back to its | |
358 | * corresponding user virtual address. | |
6a46079c AK |
359 | */ |
360 | static void add_to_kill(struct task_struct *tsk, struct page *p, | |
c36e2024 SR |
361 | pgoff_t fsdax_pgoff, struct vm_area_struct *vma, |
362 | struct list_head *to_kill) | |
6a46079c AK |
363 | { |
364 | struct to_kill *tk; | |
365 | ||
996ff7a0 JC |
366 | tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC); |
367 | if (!tk) { | |
96f96763 | 368 | pr_err("Out of memory while machine check handling\n"); |
996ff7a0 | 369 | return; |
6a46079c | 370 | } |
996ff7a0 | 371 | |
6a46079c | 372 | tk->addr = page_address_in_vma(p, vma); |
c36e2024 | 373 | if (is_zone_device_page(p)) { |
ac87ca0e | 374 | if (fsdax_pgoff != FSDAX_INVALID_PGOFF) |
c36e2024 SR |
375 | tk->addr = vma_pgoff_address(fsdax_pgoff, 1, vma); |
376 | tk->size_shift = dev_pagemap_mapping_shift(vma, tk->addr); | |
377 | } else | |
75068518 | 378 | tk->size_shift = page_shift(compound_head(p)); |
6a46079c AK |
379 | |
380 | /* | |
3d7fed4a JC |
381 | * Send SIGKILL if "tk->addr == -EFAULT". Also, as |
382 | * "tk->size_shift" is always non-zero for !is_zone_device_page(), | |
383 | * so "tk->size_shift == 0" effectively checks no mapping on | |
384 | * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times | |
385 | * to a process' address space, it's possible not all N VMAs | |
386 | * contain mappings for the page, but at least one VMA does. | |
387 | * Only deliver SIGBUS with payload derived from the VMA that | |
388 | * has a mapping for the page. | |
6a46079c | 389 | */ |
3d7fed4a | 390 | if (tk->addr == -EFAULT) { |
96f96763 | 391 | pr_info("Unable to find user space address %lx in %s\n", |
6a46079c | 392 | page_to_pfn(p), tsk->comm); |
3d7fed4a JC |
393 | } else if (tk->size_shift == 0) { |
394 | kfree(tk); | |
395 | return; | |
6a46079c | 396 | } |
996ff7a0 | 397 | |
6a46079c AK |
398 | get_task_struct(tsk); |
399 | tk->tsk = tsk; | |
400 | list_add_tail(&tk->nd, to_kill); | |
401 | } | |
402 | ||
403 | /* | |
404 | * Kill the processes that have been collected earlier. | |
405 | * | |
a21c184f ML |
406 | * Only do anything when FORCEKILL is set, otherwise just free the |
407 | * list (this is used for clean pages which do not need killing) | |
6a46079c AK |
408 | * Also when FAIL is set do a force kill because something went |
409 | * wrong earlier. | |
410 | */ | |
ae1139ec DW |
411 | static void kill_procs(struct list_head *to_kill, int forcekill, bool fail, |
412 | unsigned long pfn, int flags) | |
6a46079c AK |
413 | { |
414 | struct to_kill *tk, *next; | |
415 | ||
416 | list_for_each_entry_safe (tk, next, to_kill, nd) { | |
6751ed65 | 417 | if (forcekill) { |
6a46079c | 418 | /* |
af901ca1 | 419 | * In case something went wrong with munmapping |
6a46079c AK |
420 | * make sure the process doesn't catch the |
421 | * signal and then access the memory. Just kill it. | |
6a46079c | 422 | */ |
3d7fed4a | 423 | if (fail || tk->addr == -EFAULT) { |
96f96763 | 424 | pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n", |
1170532b | 425 | pfn, tk->tsk->comm, tk->tsk->pid); |
6376360e NH |
426 | do_send_sig_info(SIGKILL, SEND_SIG_PRIV, |
427 | tk->tsk, PIDTYPE_PID); | |
6a46079c AK |
428 | } |
429 | ||
430 | /* | |
431 | * In theory the process could have mapped | |
432 | * something else on the address in-between. We could | |
433 | * check for that, but we need to tell the | |
434 | * process anyways. | |
435 | */ | |
ae1139ec | 436 | else if (kill_proc(tk, pfn, flags) < 0) |
96f96763 | 437 | pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n", |
1170532b | 438 | pfn, tk->tsk->comm, tk->tsk->pid); |
6a46079c AK |
439 | } |
440 | put_task_struct(tk->tsk); | |
441 | kfree(tk); | |
442 | } | |
443 | } | |
444 | ||
3ba08129 NH |
445 | /* |
446 | * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO) | |
447 | * on behalf of the thread group. Return task_struct of the (first found) | |
448 | * dedicated thread if found, and return NULL otherwise. | |
449 | * | |
450 | * We already hold read_lock(&tasklist_lock) in the caller, so we don't | |
451 | * have to call rcu_read_lock/unlock() in this function. | |
452 | */ | |
453 | static struct task_struct *find_early_kill_thread(struct task_struct *tsk) | |
6a46079c | 454 | { |
3ba08129 NH |
455 | struct task_struct *t; |
456 | ||
4e018b45 NH |
457 | for_each_thread(tsk, t) { |
458 | if (t->flags & PF_MCE_PROCESS) { | |
459 | if (t->flags & PF_MCE_EARLY) | |
460 | return t; | |
461 | } else { | |
462 | if (sysctl_memory_failure_early_kill) | |
463 | return t; | |
464 | } | |
465 | } | |
3ba08129 NH |
466 | return NULL; |
467 | } | |
468 | ||
469 | /* | |
470 | * Determine whether a given process is "early kill" process which expects | |
471 | * to be signaled when some page under the process is hwpoisoned. | |
472 | * Return task_struct of the dedicated thread (main thread unless explicitly | |
30c9cf49 | 473 | * specified) if the process is "early kill" and otherwise returns NULL. |
03151c6e | 474 | * |
30c9cf49 AY |
475 | * Note that the above is true for Action Optional case. For Action Required |
476 | * case, it's only meaningful to the current thread which need to be signaled | |
477 | * with SIGBUS, this error is Action Optional for other non current | |
478 | * processes sharing the same error page,if the process is "early kill", the | |
479 | * task_struct of the dedicated thread will also be returned. | |
3ba08129 NH |
480 | */ |
481 | static struct task_struct *task_early_kill(struct task_struct *tsk, | |
482 | int force_early) | |
483 | { | |
6a46079c | 484 | if (!tsk->mm) |
3ba08129 | 485 | return NULL; |
30c9cf49 AY |
486 | /* |
487 | * Comparing ->mm here because current task might represent | |
488 | * a subthread, while tsk always points to the main thread. | |
489 | */ | |
490 | if (force_early && tsk->mm == current->mm) | |
491 | return current; | |
492 | ||
4e018b45 | 493 | return find_early_kill_thread(tsk); |
6a46079c AK |
494 | } |
495 | ||
496 | /* | |
497 | * Collect processes when the error hit an anonymous page. | |
498 | */ | |
499 | static void collect_procs_anon(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 500 | int force_early) |
6a46079c | 501 | { |
9595d769 | 502 | struct folio *folio = page_folio(page); |
6a46079c AK |
503 | struct vm_area_struct *vma; |
504 | struct task_struct *tsk; | |
505 | struct anon_vma *av; | |
bf181b9f | 506 | pgoff_t pgoff; |
6a46079c | 507 | |
6d4675e6 | 508 | av = folio_lock_anon_vma_read(folio, NULL); |
6a46079c | 509 | if (av == NULL) /* Not actually mapped anymore */ |
9b679320 PZ |
510 | return; |
511 | ||
a0f7a756 | 512 | pgoff = page_to_pgoff(page); |
9b679320 | 513 | read_lock(&tasklist_lock); |
6a46079c | 514 | for_each_process (tsk) { |
5beb4930 | 515 | struct anon_vma_chain *vmac; |
3ba08129 | 516 | struct task_struct *t = task_early_kill(tsk, force_early); |
5beb4930 | 517 | |
3ba08129 | 518 | if (!t) |
6a46079c | 519 | continue; |
bf181b9f ML |
520 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, |
521 | pgoff, pgoff) { | |
5beb4930 | 522 | vma = vmac->vma; |
6a46079c AK |
523 | if (!page_mapped_in_vma(page, vma)) |
524 | continue; | |
3ba08129 | 525 | if (vma->vm_mm == t->mm) |
ac87ca0e DW |
526 | add_to_kill(t, page, FSDAX_INVALID_PGOFF, vma, |
527 | to_kill); | |
6a46079c AK |
528 | } |
529 | } | |
6a46079c | 530 | read_unlock(&tasklist_lock); |
4fc3f1d6 | 531 | page_unlock_anon_vma_read(av); |
6a46079c AK |
532 | } |
533 | ||
534 | /* | |
535 | * Collect processes when the error hit a file mapped page. | |
536 | */ | |
537 | static void collect_procs_file(struct page *page, struct list_head *to_kill, | |
996ff7a0 | 538 | int force_early) |
6a46079c AK |
539 | { |
540 | struct vm_area_struct *vma; | |
541 | struct task_struct *tsk; | |
6a46079c | 542 | struct address_space *mapping = page->mapping; |
c43bc03d | 543 | pgoff_t pgoff; |
6a46079c | 544 | |
d28eb9c8 | 545 | i_mmap_lock_read(mapping); |
9b679320 | 546 | read_lock(&tasklist_lock); |
c43bc03d | 547 | pgoff = page_to_pgoff(page); |
6a46079c | 548 | for_each_process(tsk) { |
3ba08129 | 549 | struct task_struct *t = task_early_kill(tsk, force_early); |
6a46079c | 550 | |
3ba08129 | 551 | if (!t) |
6a46079c | 552 | continue; |
6b2dbba8 | 553 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, |
6a46079c AK |
554 | pgoff) { |
555 | /* | |
556 | * Send early kill signal to tasks where a vma covers | |
557 | * the page but the corrupted page is not necessarily | |
558 | * mapped it in its pte. | |
559 | * Assume applications who requested early kill want | |
560 | * to be informed of all such data corruptions. | |
561 | */ | |
3ba08129 | 562 | if (vma->vm_mm == t->mm) |
ac87ca0e DW |
563 | add_to_kill(t, page, FSDAX_INVALID_PGOFF, vma, |
564 | to_kill); | |
6a46079c AK |
565 | } |
566 | } | |
6a46079c | 567 | read_unlock(&tasklist_lock); |
d28eb9c8 | 568 | i_mmap_unlock_read(mapping); |
6a46079c AK |
569 | } |
570 | ||
c36e2024 SR |
571 | #ifdef CONFIG_FS_DAX |
572 | /* | |
573 | * Collect processes when the error hit a fsdax page. | |
574 | */ | |
575 | static void collect_procs_fsdax(struct page *page, | |
576 | struct address_space *mapping, pgoff_t pgoff, | |
577 | struct list_head *to_kill) | |
578 | { | |
579 | struct vm_area_struct *vma; | |
580 | struct task_struct *tsk; | |
581 | ||
582 | i_mmap_lock_read(mapping); | |
583 | read_lock(&tasklist_lock); | |
584 | for_each_process(tsk) { | |
585 | struct task_struct *t = task_early_kill(tsk, true); | |
586 | ||
587 | if (!t) | |
588 | continue; | |
589 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { | |
590 | if (vma->vm_mm == t->mm) | |
591 | add_to_kill(t, page, pgoff, vma, to_kill); | |
592 | } | |
593 | } | |
594 | read_unlock(&tasklist_lock); | |
595 | i_mmap_unlock_read(mapping); | |
596 | } | |
597 | #endif /* CONFIG_FS_DAX */ | |
598 | ||
6a46079c AK |
599 | /* |
600 | * Collect the processes who have the corrupted page mapped to kill. | |
6a46079c | 601 | */ |
74614de1 TL |
602 | static void collect_procs(struct page *page, struct list_head *tokill, |
603 | int force_early) | |
6a46079c | 604 | { |
6a46079c AK |
605 | if (!page->mapping) |
606 | return; | |
607 | ||
6a46079c | 608 | if (PageAnon(page)) |
996ff7a0 | 609 | collect_procs_anon(page, tokill, force_early); |
6a46079c | 610 | else |
996ff7a0 | 611 | collect_procs_file(page, tokill, force_early); |
6a46079c AK |
612 | } |
613 | ||
a3f5d80e NH |
614 | struct hwp_walk { |
615 | struct to_kill tk; | |
616 | unsigned long pfn; | |
617 | int flags; | |
618 | }; | |
619 | ||
620 | static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift) | |
621 | { | |
622 | tk->addr = addr; | |
623 | tk->size_shift = shift; | |
624 | } | |
625 | ||
626 | static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift, | |
627 | unsigned long poisoned_pfn, struct to_kill *tk) | |
628 | { | |
629 | unsigned long pfn = 0; | |
630 | ||
631 | if (pte_present(pte)) { | |
632 | pfn = pte_pfn(pte); | |
633 | } else { | |
634 | swp_entry_t swp = pte_to_swp_entry(pte); | |
635 | ||
636 | if (is_hwpoison_entry(swp)) | |
637 | pfn = hwpoison_entry_to_pfn(swp); | |
638 | } | |
639 | ||
640 | if (!pfn || pfn != poisoned_pfn) | |
641 | return 0; | |
642 | ||
643 | set_to_kill(tk, addr, shift); | |
644 | return 1; | |
645 | } | |
646 | ||
647 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
648 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
649 | struct hwp_walk *hwp) | |
650 | { | |
651 | pmd_t pmd = *pmdp; | |
652 | unsigned long pfn; | |
653 | unsigned long hwpoison_vaddr; | |
654 | ||
655 | if (!pmd_present(pmd)) | |
656 | return 0; | |
657 | pfn = pmd_pfn(pmd); | |
658 | if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) { | |
659 | hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT); | |
660 | set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT); | |
661 | return 1; | |
662 | } | |
663 | return 0; | |
664 | } | |
665 | #else | |
666 | static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr, | |
667 | struct hwp_walk *hwp) | |
668 | { | |
669 | return 0; | |
670 | } | |
671 | #endif | |
672 | ||
673 | static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr, | |
674 | unsigned long end, struct mm_walk *walk) | |
675 | { | |
f142e707 | 676 | struct hwp_walk *hwp = walk->private; |
a3f5d80e | 677 | int ret = 0; |
ea3732f7 | 678 | pte_t *ptep, *mapped_pte; |
a3f5d80e NH |
679 | spinlock_t *ptl; |
680 | ||
681 | ptl = pmd_trans_huge_lock(pmdp, walk->vma); | |
682 | if (ptl) { | |
683 | ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp); | |
684 | spin_unlock(ptl); | |
685 | goto out; | |
686 | } | |
687 | ||
688 | if (pmd_trans_unstable(pmdp)) | |
689 | goto out; | |
690 | ||
ea3732f7 ML |
691 | mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, |
692 | addr, &ptl); | |
a3f5d80e NH |
693 | for (; addr != end; ptep++, addr += PAGE_SIZE) { |
694 | ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT, | |
695 | hwp->pfn, &hwp->tk); | |
696 | if (ret == 1) | |
697 | break; | |
698 | } | |
ea3732f7 | 699 | pte_unmap_unlock(mapped_pte, ptl); |
a3f5d80e NH |
700 | out: |
701 | cond_resched(); | |
702 | return ret; | |
703 | } | |
704 | ||
705 | #ifdef CONFIG_HUGETLB_PAGE | |
706 | static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask, | |
707 | unsigned long addr, unsigned long end, | |
708 | struct mm_walk *walk) | |
709 | { | |
f142e707 | 710 | struct hwp_walk *hwp = walk->private; |
a3f5d80e NH |
711 | pte_t pte = huge_ptep_get(ptep); |
712 | struct hstate *h = hstate_vma(walk->vma); | |
713 | ||
714 | return check_hwpoisoned_entry(pte, addr, huge_page_shift(h), | |
715 | hwp->pfn, &hwp->tk); | |
716 | } | |
717 | #else | |
718 | #define hwpoison_hugetlb_range NULL | |
719 | #endif | |
720 | ||
ba9eb3ce | 721 | static const struct mm_walk_ops hwp_walk_ops = { |
a3f5d80e NH |
722 | .pmd_entry = hwpoison_pte_range, |
723 | .hugetlb_entry = hwpoison_hugetlb_range, | |
724 | }; | |
725 | ||
726 | /* | |
727 | * Sends SIGBUS to the current process with error info. | |
728 | * | |
729 | * This function is intended to handle "Action Required" MCEs on already | |
730 | * hardware poisoned pages. They could happen, for example, when | |
731 | * memory_failure() failed to unmap the error page at the first call, or | |
732 | * when multiple local machine checks happened on different CPUs. | |
733 | * | |
734 | * MCE handler currently has no easy access to the error virtual address, | |
735 | * so this function walks page table to find it. The returned virtual address | |
736 | * is proper in most cases, but it could be wrong when the application | |
737 | * process has multiple entries mapping the error page. | |
738 | */ | |
739 | static int kill_accessing_process(struct task_struct *p, unsigned long pfn, | |
740 | int flags) | |
741 | { | |
742 | int ret; | |
743 | struct hwp_walk priv = { | |
744 | .pfn = pfn, | |
745 | }; | |
746 | priv.tk.tsk = p; | |
747 | ||
77677cdb SX |
748 | if (!p->mm) |
749 | return -EFAULT; | |
750 | ||
a3f5d80e NH |
751 | mmap_read_lock(p->mm); |
752 | ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops, | |
753 | (void *)&priv); | |
754 | if (ret == 1 && priv.tk.addr) | |
755 | kill_proc(&priv.tk, pfn, flags); | |
046545a6 NH |
756 | else |
757 | ret = 0; | |
a3f5d80e | 758 | mmap_read_unlock(p->mm); |
046545a6 | 759 | return ret > 0 ? -EHWPOISON : -EFAULT; |
a3f5d80e NH |
760 | } |
761 | ||
6a46079c | 762 | static const char *action_name[] = { |
cc637b17 XX |
763 | [MF_IGNORED] = "Ignored", |
764 | [MF_FAILED] = "Failed", | |
765 | [MF_DELAYED] = "Delayed", | |
766 | [MF_RECOVERED] = "Recovered", | |
64d37a2b NH |
767 | }; |
768 | ||
769 | static const char * const action_page_types[] = { | |
cc637b17 XX |
770 | [MF_MSG_KERNEL] = "reserved kernel page", |
771 | [MF_MSG_KERNEL_HIGH_ORDER] = "high-order kernel page", | |
772 | [MF_MSG_SLAB] = "kernel slab page", | |
773 | [MF_MSG_DIFFERENT_COMPOUND] = "different compound page after locking", | |
cc637b17 XX |
774 | [MF_MSG_HUGE] = "huge page", |
775 | [MF_MSG_FREE_HUGE] = "free huge page", | |
776 | [MF_MSG_UNMAP_FAILED] = "unmapping failed page", | |
777 | [MF_MSG_DIRTY_SWAPCACHE] = "dirty swapcache page", | |
778 | [MF_MSG_CLEAN_SWAPCACHE] = "clean swapcache page", | |
779 | [MF_MSG_DIRTY_MLOCKED_LRU] = "dirty mlocked LRU page", | |
780 | [MF_MSG_CLEAN_MLOCKED_LRU] = "clean mlocked LRU page", | |
781 | [MF_MSG_DIRTY_UNEVICTABLE_LRU] = "dirty unevictable LRU page", | |
782 | [MF_MSG_CLEAN_UNEVICTABLE_LRU] = "clean unevictable LRU page", | |
783 | [MF_MSG_DIRTY_LRU] = "dirty LRU page", | |
784 | [MF_MSG_CLEAN_LRU] = "clean LRU page", | |
785 | [MF_MSG_TRUNCATED_LRU] = "already truncated LRU page", | |
786 | [MF_MSG_BUDDY] = "free buddy page", | |
6100e34b | 787 | [MF_MSG_DAX] = "dax page", |
5d1fd5dc | 788 | [MF_MSG_UNSPLIT_THP] = "unsplit thp", |
cc637b17 | 789 | [MF_MSG_UNKNOWN] = "unknown page", |
64d37a2b NH |
790 | }; |
791 | ||
dc2a1cbf WF |
792 | /* |
793 | * XXX: It is possible that a page is isolated from LRU cache, | |
794 | * and then kept in swap cache or failed to remove from page cache. | |
795 | * The page count will stop it from being freed by unpoison. | |
796 | * Stress tests should be aware of this memory leak problem. | |
797 | */ | |
798 | static int delete_from_lru_cache(struct page *p) | |
799 | { | |
800 | if (!isolate_lru_page(p)) { | |
801 | /* | |
802 | * Clear sensible page flags, so that the buddy system won't | |
803 | * complain when the page is unpoison-and-freed. | |
804 | */ | |
805 | ClearPageActive(p); | |
806 | ClearPageUnevictable(p); | |
18365225 MH |
807 | |
808 | /* | |
809 | * Poisoned page might never drop its ref count to 0 so we have | |
810 | * to uncharge it manually from its memcg. | |
811 | */ | |
bbc6b703 | 812 | mem_cgroup_uncharge(page_folio(p)); |
18365225 | 813 | |
dc2a1cbf WF |
814 | /* |
815 | * drop the page count elevated by isolate_lru_page() | |
816 | */ | |
09cbfeaf | 817 | put_page(p); |
dc2a1cbf WF |
818 | return 0; |
819 | } | |
820 | return -EIO; | |
821 | } | |
822 | ||
78bb9203 NH |
823 | static int truncate_error_page(struct page *p, unsigned long pfn, |
824 | struct address_space *mapping) | |
825 | { | |
826 | int ret = MF_FAILED; | |
827 | ||
828 | if (mapping->a_ops->error_remove_page) { | |
829 | int err = mapping->a_ops->error_remove_page(mapping, p); | |
830 | ||
831 | if (err != 0) { | |
96f96763 | 832 | pr_info("%#lx: Failed to punch page: %d\n", pfn, err); |
78bb9203 NH |
833 | } else if (page_has_private(p) && |
834 | !try_to_release_page(p, GFP_NOIO)) { | |
96f96763 | 835 | pr_info("%#lx: failed to release buffers\n", pfn); |
78bb9203 NH |
836 | } else { |
837 | ret = MF_RECOVERED; | |
838 | } | |
839 | } else { | |
840 | /* | |
841 | * If the file system doesn't support it just invalidate | |
842 | * This fails on dirty or anything with private pages | |
843 | */ | |
844 | if (invalidate_inode_page(p)) | |
845 | ret = MF_RECOVERED; | |
846 | else | |
96f96763 | 847 | pr_info("%#lx: Failed to invalidate\n", pfn); |
78bb9203 NH |
848 | } |
849 | ||
850 | return ret; | |
851 | } | |
852 | ||
dd0f230a YS |
853 | struct page_state { |
854 | unsigned long mask; | |
855 | unsigned long res; | |
856 | enum mf_action_page_type type; | |
857 | ||
858 | /* Callback ->action() has to unlock the relevant page inside it. */ | |
859 | int (*action)(struct page_state *ps, struct page *p); | |
860 | }; | |
861 | ||
862 | /* | |
863 | * Return true if page is still referenced by others, otherwise return | |
864 | * false. | |
865 | * | |
866 | * The extra_pins is true when one extra refcount is expected. | |
867 | */ | |
868 | static bool has_extra_refcount(struct page_state *ps, struct page *p, | |
869 | bool extra_pins) | |
870 | { | |
871 | int count = page_count(p) - 1; | |
872 | ||
873 | if (extra_pins) | |
874 | count -= 1; | |
875 | ||
876 | if (count > 0) { | |
96f96763 | 877 | pr_err("%#lx: %s still referenced by %d users\n", |
dd0f230a YS |
878 | page_to_pfn(p), action_page_types[ps->type], count); |
879 | return true; | |
880 | } | |
881 | ||
882 | return false; | |
883 | } | |
884 | ||
6a46079c AK |
885 | /* |
886 | * Error hit kernel page. | |
887 | * Do nothing, try to be lucky and not touch this instead. For a few cases we | |
888 | * could be more sophisticated. | |
889 | */ | |
dd0f230a | 890 | static int me_kernel(struct page_state *ps, struct page *p) |
6a46079c | 891 | { |
ea6d0630 | 892 | unlock_page(p); |
cc637b17 | 893 | return MF_IGNORED; |
6a46079c AK |
894 | } |
895 | ||
896 | /* | |
897 | * Page in unknown state. Do nothing. | |
898 | */ | |
dd0f230a | 899 | static int me_unknown(struct page_state *ps, struct page *p) |
6a46079c | 900 | { |
96f96763 | 901 | pr_err("%#lx: Unknown page state\n", page_to_pfn(p)); |
ea6d0630 | 902 | unlock_page(p); |
cc637b17 | 903 | return MF_FAILED; |
6a46079c AK |
904 | } |
905 | ||
6a46079c AK |
906 | /* |
907 | * Clean (or cleaned) page cache page. | |
908 | */ | |
dd0f230a | 909 | static int me_pagecache_clean(struct page_state *ps, struct page *p) |
6a46079c | 910 | { |
ea6d0630 | 911 | int ret; |
6a46079c | 912 | struct address_space *mapping; |
a7605426 | 913 | bool extra_pins; |
6a46079c | 914 | |
dc2a1cbf WF |
915 | delete_from_lru_cache(p); |
916 | ||
6a46079c AK |
917 | /* |
918 | * For anonymous pages we're done the only reference left | |
919 | * should be the one m_f() holds. | |
920 | */ | |
ea6d0630 NH |
921 | if (PageAnon(p)) { |
922 | ret = MF_RECOVERED; | |
923 | goto out; | |
924 | } | |
6a46079c AK |
925 | |
926 | /* | |
927 | * Now truncate the page in the page cache. This is really | |
928 | * more like a "temporary hole punch" | |
929 | * Don't do this for block devices when someone else | |
930 | * has a reference, because it could be file system metadata | |
931 | * and that's not safe to truncate. | |
932 | */ | |
933 | mapping = page_mapping(p); | |
934 | if (!mapping) { | |
935 | /* | |
936 | * Page has been teared down in the meanwhile | |
937 | */ | |
ea6d0630 NH |
938 | ret = MF_FAILED; |
939 | goto out; | |
6a46079c AK |
940 | } |
941 | ||
a7605426 YS |
942 | /* |
943 | * The shmem page is kept in page cache instead of truncating | |
944 | * so is expected to have an extra refcount after error-handling. | |
945 | */ | |
946 | extra_pins = shmem_mapping(mapping); | |
947 | ||
6a46079c AK |
948 | /* |
949 | * Truncation is a bit tricky. Enable it per file system for now. | |
950 | * | |
9608703e | 951 | * Open: to take i_rwsem or not for this? Right now we don't. |
6a46079c | 952 | */ |
dd0f230a | 953 | ret = truncate_error_page(p, page_to_pfn(p), mapping); |
a7605426 YS |
954 | if (has_extra_refcount(ps, p, extra_pins)) |
955 | ret = MF_FAILED; | |
956 | ||
ea6d0630 NH |
957 | out: |
958 | unlock_page(p); | |
dd0f230a | 959 | |
ea6d0630 | 960 | return ret; |
6a46079c AK |
961 | } |
962 | ||
963 | /* | |
549543df | 964 | * Dirty pagecache page |
6a46079c AK |
965 | * Issues: when the error hit a hole page the error is not properly |
966 | * propagated. | |
967 | */ | |
dd0f230a | 968 | static int me_pagecache_dirty(struct page_state *ps, struct page *p) |
6a46079c AK |
969 | { |
970 | struct address_space *mapping = page_mapping(p); | |
971 | ||
972 | SetPageError(p); | |
973 | /* TBD: print more information about the file. */ | |
974 | if (mapping) { | |
975 | /* | |
976 | * IO error will be reported by write(), fsync(), etc. | |
977 | * who check the mapping. | |
978 | * This way the application knows that something went | |
979 | * wrong with its dirty file data. | |
980 | * | |
981 | * There's one open issue: | |
982 | * | |
983 | * The EIO will be only reported on the next IO | |
984 | * operation and then cleared through the IO map. | |
985 | * Normally Linux has two mechanisms to pass IO error | |
986 | * first through the AS_EIO flag in the address space | |
987 | * and then through the PageError flag in the page. | |
988 | * Since we drop pages on memory failure handling the | |
989 | * only mechanism open to use is through AS_AIO. | |
990 | * | |
991 | * This has the disadvantage that it gets cleared on | |
992 | * the first operation that returns an error, while | |
993 | * the PageError bit is more sticky and only cleared | |
994 | * when the page is reread or dropped. If an | |
995 | * application assumes it will always get error on | |
996 | * fsync, but does other operations on the fd before | |
25985edc | 997 | * and the page is dropped between then the error |
6a46079c AK |
998 | * will not be properly reported. |
999 | * | |
1000 | * This can already happen even without hwpoisoned | |
1001 | * pages: first on metadata IO errors (which only | |
1002 | * report through AS_EIO) or when the page is dropped | |
1003 | * at the wrong time. | |
1004 | * | |
1005 | * So right now we assume that the application DTRT on | |
1006 | * the first EIO, but we're not worse than other parts | |
1007 | * of the kernel. | |
1008 | */ | |
af21bfaf | 1009 | mapping_set_error(mapping, -EIO); |
6a46079c AK |
1010 | } |
1011 | ||
dd0f230a | 1012 | return me_pagecache_clean(ps, p); |
6a46079c AK |
1013 | } |
1014 | ||
1015 | /* | |
1016 | * Clean and dirty swap cache. | |
1017 | * | |
1018 | * Dirty swap cache page is tricky to handle. The page could live both in page | |
1019 | * cache and swap cache(ie. page is freshly swapped in). So it could be | |
1020 | * referenced concurrently by 2 types of PTEs: | |
1021 | * normal PTEs and swap PTEs. We try to handle them consistently by calling | |
1022 | * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs, | |
1023 | * and then | |
1024 | * - clear dirty bit to prevent IO | |
1025 | * - remove from LRU | |
1026 | * - but keep in the swap cache, so that when we return to it on | |
1027 | * a later page fault, we know the application is accessing | |
1028 | * corrupted data and shall be killed (we installed simple | |
1029 | * interception code in do_swap_page to catch it). | |
1030 | * | |
1031 | * Clean swap cache pages can be directly isolated. A later page fault will | |
1032 | * bring in the known good data from disk. | |
1033 | */ | |
dd0f230a | 1034 | static int me_swapcache_dirty(struct page_state *ps, struct page *p) |
6a46079c | 1035 | { |
ea6d0630 | 1036 | int ret; |
dd0f230a | 1037 | bool extra_pins = false; |
ea6d0630 | 1038 | |
6a46079c AK |
1039 | ClearPageDirty(p); |
1040 | /* Trigger EIO in shmem: */ | |
1041 | ClearPageUptodate(p); | |
1042 | ||
ea6d0630 NH |
1043 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED; |
1044 | unlock_page(p); | |
dd0f230a YS |
1045 | |
1046 | if (ret == MF_DELAYED) | |
1047 | extra_pins = true; | |
1048 | ||
1049 | if (has_extra_refcount(ps, p, extra_pins)) | |
1050 | ret = MF_FAILED; | |
1051 | ||
ea6d0630 | 1052 | return ret; |
6a46079c AK |
1053 | } |
1054 | ||
dd0f230a | 1055 | static int me_swapcache_clean(struct page_state *ps, struct page *p) |
6a46079c | 1056 | { |
75fa68a5 | 1057 | struct folio *folio = page_folio(p); |
ea6d0630 NH |
1058 | int ret; |
1059 | ||
75fa68a5 | 1060 | delete_from_swap_cache(folio); |
e43c3afb | 1061 | |
ea6d0630 | 1062 | ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED; |
75fa68a5 | 1063 | folio_unlock(folio); |
dd0f230a YS |
1064 | |
1065 | if (has_extra_refcount(ps, p, false)) | |
1066 | ret = MF_FAILED; | |
1067 | ||
ea6d0630 | 1068 | return ret; |
6a46079c AK |
1069 | } |
1070 | ||
1071 | /* | |
1072 | * Huge pages. Needs work. | |
1073 | * Issues: | |
93f70f90 NH |
1074 | * - Error on hugepage is contained in hugepage unit (not in raw page unit.) |
1075 | * To narrow down kill region to one page, we need to break up pmd. | |
6a46079c | 1076 | */ |
dd0f230a | 1077 | static int me_huge_page(struct page_state *ps, struct page *p) |
6a46079c | 1078 | { |
a8b2c2ce | 1079 | int res; |
93f70f90 | 1080 | struct page *hpage = compound_head(p); |
78bb9203 | 1081 | struct address_space *mapping; |
2491ffee NH |
1082 | |
1083 | if (!PageHuge(hpage)) | |
1084 | return MF_DELAYED; | |
1085 | ||
78bb9203 NH |
1086 | mapping = page_mapping(hpage); |
1087 | if (mapping) { | |
dd0f230a | 1088 | res = truncate_error_page(hpage, page_to_pfn(p), mapping); |
ea6d0630 | 1089 | unlock_page(hpage); |
78bb9203 NH |
1090 | } else { |
1091 | unlock_page(hpage); | |
1092 | /* | |
ef526b17 ML |
1093 | * migration entry prevents later access on error hugepage, |
1094 | * so we can free and dissolve it into buddy to save healthy | |
1095 | * subpages. | |
78bb9203 | 1096 | */ |
ef526b17 | 1097 | put_page(hpage); |
ceaf8fbe | 1098 | if (__page_handle_poison(p) >= 0) { |
a8b2c2ce OS |
1099 | page_ref_inc(p); |
1100 | res = MF_RECOVERED; | |
ceaf8fbe NH |
1101 | } else { |
1102 | res = MF_FAILED; | |
a8b2c2ce | 1103 | } |
93f70f90 | 1104 | } |
78bb9203 | 1105 | |
dd0f230a YS |
1106 | if (has_extra_refcount(ps, p, false)) |
1107 | res = MF_FAILED; | |
1108 | ||
78bb9203 | 1109 | return res; |
6a46079c AK |
1110 | } |
1111 | ||
1112 | /* | |
1113 | * Various page states we can handle. | |
1114 | * | |
1115 | * A page state is defined by its current page->flags bits. | |
1116 | * The table matches them in order and calls the right handler. | |
1117 | * | |
1118 | * This is quite tricky because we can access page at any time | |
25985edc | 1119 | * in its live cycle, so all accesses have to be extremely careful. |
6a46079c AK |
1120 | * |
1121 | * This is not complete. More states could be added. | |
1122 | * For any missing state don't attempt recovery. | |
1123 | */ | |
1124 | ||
1125 | #define dirty (1UL << PG_dirty) | |
6326fec1 | 1126 | #define sc ((1UL << PG_swapcache) | (1UL << PG_swapbacked)) |
6a46079c AK |
1127 | #define unevict (1UL << PG_unevictable) |
1128 | #define mlock (1UL << PG_mlocked) | |
6a46079c | 1129 | #define lru (1UL << PG_lru) |
6a46079c | 1130 | #define head (1UL << PG_head) |
6a46079c | 1131 | #define slab (1UL << PG_slab) |
6a46079c AK |
1132 | #define reserved (1UL << PG_reserved) |
1133 | ||
dd0f230a | 1134 | static struct page_state error_states[] = { |
cc637b17 | 1135 | { reserved, reserved, MF_MSG_KERNEL, me_kernel }, |
95d01fc6 WF |
1136 | /* |
1137 | * free pages are specially detected outside this table: | |
1138 | * PG_buddy pages only make a small fraction of all free pages. | |
1139 | */ | |
6a46079c AK |
1140 | |
1141 | /* | |
1142 | * Could in theory check if slab page is free or if we can drop | |
1143 | * currently unused objects without touching them. But just | |
1144 | * treat it as standard kernel for now. | |
1145 | */ | |
cc637b17 | 1146 | { slab, slab, MF_MSG_SLAB, me_kernel }, |
6a46079c | 1147 | |
cc637b17 | 1148 | { head, head, MF_MSG_HUGE, me_huge_page }, |
6a46079c | 1149 | |
cc637b17 XX |
1150 | { sc|dirty, sc|dirty, MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty }, |
1151 | { sc|dirty, sc, MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean }, | |
6a46079c | 1152 | |
cc637b17 XX |
1153 | { mlock|dirty, mlock|dirty, MF_MSG_DIRTY_MLOCKED_LRU, me_pagecache_dirty }, |
1154 | { mlock|dirty, mlock, MF_MSG_CLEAN_MLOCKED_LRU, me_pagecache_clean }, | |
6a46079c | 1155 | |
cc637b17 XX |
1156 | { unevict|dirty, unevict|dirty, MF_MSG_DIRTY_UNEVICTABLE_LRU, me_pagecache_dirty }, |
1157 | { unevict|dirty, unevict, MF_MSG_CLEAN_UNEVICTABLE_LRU, me_pagecache_clean }, | |
5f4b9fc5 | 1158 | |
cc637b17 XX |
1159 | { lru|dirty, lru|dirty, MF_MSG_DIRTY_LRU, me_pagecache_dirty }, |
1160 | { lru|dirty, lru, MF_MSG_CLEAN_LRU, me_pagecache_clean }, | |
6a46079c AK |
1161 | |
1162 | /* | |
1163 | * Catchall entry: must be at end. | |
1164 | */ | |
cc637b17 | 1165 | { 0, 0, MF_MSG_UNKNOWN, me_unknown }, |
6a46079c AK |
1166 | }; |
1167 | ||
2326c467 AK |
1168 | #undef dirty |
1169 | #undef sc | |
1170 | #undef unevict | |
1171 | #undef mlock | |
2326c467 | 1172 | #undef lru |
2326c467 | 1173 | #undef head |
2326c467 AK |
1174 | #undef slab |
1175 | #undef reserved | |
1176 | ||
ff604cf6 NH |
1177 | /* |
1178 | * "Dirty/Clean" indication is not 100% accurate due to the possibility of | |
1179 | * setting PG_dirty outside page lock. See also comment above set_page_dirty(). | |
1180 | */ | |
cc3e2af4 XX |
1181 | static void action_result(unsigned long pfn, enum mf_action_page_type type, |
1182 | enum mf_result result) | |
6a46079c | 1183 | { |
97f0b134 XX |
1184 | trace_memory_failure_event(pfn, type, result); |
1185 | ||
e240ac52 | 1186 | num_poisoned_pages_inc(); |
96f96763 | 1187 | pr_err("%#lx: recovery action for %s: %s\n", |
64d37a2b | 1188 | pfn, action_page_types[type], action_name[result]); |
6a46079c AK |
1189 | } |
1190 | ||
1191 | static int page_action(struct page_state *ps, struct page *p, | |
bd1ce5f9 | 1192 | unsigned long pfn) |
6a46079c AK |
1193 | { |
1194 | int result; | |
1195 | ||
ea6d0630 | 1196 | /* page p should be unlocked after returning from ps->action(). */ |
dd0f230a | 1197 | result = ps->action(ps, p); |
7456b040 | 1198 | |
64d37a2b | 1199 | action_result(pfn, ps->type, result); |
6a46079c AK |
1200 | |
1201 | /* Could do more checks here if page looks ok */ | |
1202 | /* | |
1203 | * Could adjust zone counters here to correct for the missing page. | |
1204 | */ | |
1205 | ||
cc637b17 | 1206 | return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY; |
6a46079c AK |
1207 | } |
1208 | ||
bf181c58 NH |
1209 | static inline bool PageHWPoisonTakenOff(struct page *page) |
1210 | { | |
1211 | return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON; | |
1212 | } | |
1213 | ||
1214 | void SetPageHWPoisonTakenOff(struct page *page) | |
1215 | { | |
1216 | set_page_private(page, MAGIC_HWPOISON); | |
1217 | } | |
1218 | ||
1219 | void ClearPageHWPoisonTakenOff(struct page *page) | |
1220 | { | |
1221 | if (PageHWPoison(page)) | |
1222 | set_page_private(page, 0); | |
1223 | } | |
1224 | ||
25182f05 NH |
1225 | /* |
1226 | * Return true if a page type of a given page is supported by hwpoison | |
1227 | * mechanism (while handling could fail), otherwise false. This function | |
1228 | * does not return true for hugetlb or device memory pages, so it's assumed | |
1229 | * to be called only in the context where we never have such pages. | |
1230 | */ | |
bf6445bc | 1231 | static inline bool HWPoisonHandlable(struct page *page, unsigned long flags) |
25182f05 | 1232 | { |
3f871370 | 1233 | /* Soft offline could migrate non-LRU movable pages */ |
bf6445bc | 1234 | if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page)) |
3f871370 | 1235 | return true; |
bf6445bc | 1236 | |
3f871370 | 1237 | return PageLRU(page) || is_free_buddy_page(page); |
25182f05 NH |
1238 | } |
1239 | ||
bf6445bc | 1240 | static int __get_hwpoison_page(struct page *page, unsigned long flags) |
ead07f6a NH |
1241 | { |
1242 | struct page *head = compound_head(page); | |
25182f05 NH |
1243 | int ret = 0; |
1244 | bool hugetlb = false; | |
1245 | ||
1246 | ret = get_hwpoison_huge_page(head, &hugetlb); | |
1247 | if (hugetlb) | |
1248 | return ret; | |
1249 | ||
1250 | /* | |
1251 | * This check prevents from calling get_hwpoison_unless_zero() | |
1252 | * for any unsupported type of page in order to reduce the risk of | |
1253 | * unexpected races caused by taking a page refcount. | |
1254 | */ | |
bf6445bc | 1255 | if (!HWPoisonHandlable(head, flags)) |
fcc00621 | 1256 | return -EBUSY; |
ead07f6a | 1257 | |
c2e7e00b KK |
1258 | if (get_page_unless_zero(head)) { |
1259 | if (head == compound_head(page)) | |
1260 | return 1; | |
1261 | ||
96f96763 | 1262 | pr_info("%#lx cannot catch tail\n", page_to_pfn(page)); |
c2e7e00b KK |
1263 | put_page(head); |
1264 | } | |
1265 | ||
1266 | return 0; | |
ead07f6a | 1267 | } |
ead07f6a | 1268 | |
2f714160 | 1269 | static int get_any_page(struct page *p, unsigned long flags) |
17e395b6 | 1270 | { |
2f714160 OS |
1271 | int ret = 0, pass = 0; |
1272 | bool count_increased = false; | |
17e395b6 | 1273 | |
2f714160 OS |
1274 | if (flags & MF_COUNT_INCREASED) |
1275 | count_increased = true; | |
1276 | ||
1277 | try_again: | |
0ed950d1 | 1278 | if (!count_increased) { |
bf6445bc | 1279 | ret = __get_hwpoison_page(p, flags); |
0ed950d1 NH |
1280 | if (!ret) { |
1281 | if (page_count(p)) { | |
1282 | /* We raced with an allocation, retry. */ | |
1283 | if (pass++ < 3) | |
1284 | goto try_again; | |
1285 | ret = -EBUSY; | |
1286 | } else if (!PageHuge(p) && !is_free_buddy_page(p)) { | |
1287 | /* We raced with put_page, retry. */ | |
1288 | if (pass++ < 3) | |
1289 | goto try_again; | |
1290 | ret = -EIO; | |
1291 | } | |
1292 | goto out; | |
1293 | } else if (ret == -EBUSY) { | |
fcc00621 NH |
1294 | /* |
1295 | * We raced with (possibly temporary) unhandlable | |
1296 | * page, retry. | |
1297 | */ | |
1298 | if (pass++ < 3) { | |
d0505e9f | 1299 | shake_page(p); |
2f714160 | 1300 | goto try_again; |
fcc00621 NH |
1301 | } |
1302 | ret = -EIO; | |
0ed950d1 | 1303 | goto out; |
2f714160 | 1304 | } |
0ed950d1 NH |
1305 | } |
1306 | ||
bf6445bc | 1307 | if (PageHuge(p) || HWPoisonHandlable(p, flags)) { |
0ed950d1 | 1308 | ret = 1; |
2f714160 | 1309 | } else { |
0ed950d1 NH |
1310 | /* |
1311 | * A page we cannot handle. Check whether we can turn | |
1312 | * it into something we can handle. | |
1313 | */ | |
1314 | if (pass++ < 3) { | |
2f714160 | 1315 | put_page(p); |
d0505e9f | 1316 | shake_page(p); |
0ed950d1 NH |
1317 | count_increased = false; |
1318 | goto try_again; | |
2f714160 | 1319 | } |
0ed950d1 NH |
1320 | put_page(p); |
1321 | ret = -EIO; | |
17e395b6 | 1322 | } |
0ed950d1 | 1323 | out: |
941ca063 | 1324 | if (ret == -EIO) |
96f96763 | 1325 | pr_err("%#lx: unhandlable page.\n", page_to_pfn(p)); |
941ca063 | 1326 | |
17e395b6 OS |
1327 | return ret; |
1328 | } | |
1329 | ||
bf181c58 NH |
1330 | static int __get_unpoison_page(struct page *page) |
1331 | { | |
1332 | struct page *head = compound_head(page); | |
1333 | int ret = 0; | |
1334 | bool hugetlb = false; | |
1335 | ||
1336 | ret = get_hwpoison_huge_page(head, &hugetlb); | |
1337 | if (hugetlb) | |
1338 | return ret; | |
1339 | ||
1340 | /* | |
1341 | * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison, | |
1342 | * but also isolated from buddy freelist, so need to identify the | |
1343 | * state and have to cancel both operations to unpoison. | |
1344 | */ | |
1345 | if (PageHWPoisonTakenOff(page)) | |
1346 | return -EHWPOISON; | |
1347 | ||
1348 | return get_page_unless_zero(page) ? 1 : 0; | |
1349 | } | |
1350 | ||
0ed950d1 NH |
1351 | /** |
1352 | * get_hwpoison_page() - Get refcount for memory error handling | |
1353 | * @p: Raw error page (hit by memory error) | |
1354 | * @flags: Flags controlling behavior of error handling | |
1355 | * | |
1356 | * get_hwpoison_page() takes a page refcount of an error page to handle memory | |
1357 | * error on it, after checking that the error page is in a well-defined state | |
0b8f0d87 | 1358 | * (defined as a page-type we can successfully handle the memory error on it, |
0ed950d1 NH |
1359 | * such as LRU page and hugetlb page). |
1360 | * | |
1361 | * Memory error handling could be triggered at any time on any type of page, | |
1362 | * so it's prone to race with typical memory management lifecycle (like | |
1363 | * allocation and free). So to avoid such races, get_hwpoison_page() takes | |
1364 | * extra care for the error page's state (as done in __get_hwpoison_page()), | |
1365 | * and has some retry logic in get_any_page(). | |
1366 | * | |
bf181c58 NH |
1367 | * When called from unpoison_memory(), the caller should already ensure that |
1368 | * the given page has PG_hwpoison. So it's never reused for other page | |
1369 | * allocations, and __get_unpoison_page() never races with them. | |
1370 | * | |
0ed950d1 NH |
1371 | * Return: 0 on failure, |
1372 | * 1 on success for in-use pages in a well-defined state, | |
1373 | * -EIO for pages on which we can not handle memory errors, | |
1374 | * -EBUSY when get_hwpoison_page() has raced with page lifecycle | |
bf181c58 NH |
1375 | * operations like allocation and free, |
1376 | * -EHWPOISON when the page is hwpoisoned and taken off from buddy. | |
0ed950d1 NH |
1377 | */ |
1378 | static int get_hwpoison_page(struct page *p, unsigned long flags) | |
2f714160 OS |
1379 | { |
1380 | int ret; | |
1381 | ||
1382 | zone_pcp_disable(page_zone(p)); | |
bf181c58 NH |
1383 | if (flags & MF_UNPOISON) |
1384 | ret = __get_unpoison_page(p); | |
1385 | else | |
1386 | ret = get_any_page(p, flags); | |
2f714160 OS |
1387 | zone_pcp_enable(page_zone(p)); |
1388 | ||
1389 | return ret; | |
1390 | } | |
1391 | ||
6a46079c AK |
1392 | /* |
1393 | * Do all that is necessary to remove user space mappings. Unmap | |
1394 | * the pages and send SIGBUS to the processes if the data was dirty. | |
1395 | */ | |
666e5a40 | 1396 | static bool hwpoison_user_mappings(struct page *p, unsigned long pfn, |
ed8c2f49 | 1397 | int flags, struct page *hpage) |
6a46079c | 1398 | { |
869f7ee6 | 1399 | struct folio *folio = page_folio(hpage); |
36af6737 | 1400 | enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC; |
6a46079c AK |
1401 | struct address_space *mapping; |
1402 | LIST_HEAD(tokill); | |
1fb08ac6 | 1403 | bool unmap_success; |
6751ed65 | 1404 | int kill = 1, forcekill; |
286c469a | 1405 | bool mlocked = PageMlocked(hpage); |
6a46079c | 1406 | |
93a9eb39 NH |
1407 | /* |
1408 | * Here we are interested only in user-mapped pages, so skip any | |
1409 | * other types of pages. | |
1410 | */ | |
1411 | if (PageReserved(p) || PageSlab(p)) | |
666e5a40 | 1412 | return true; |
93a9eb39 | 1413 | if (!(PageLRU(hpage) || PageHuge(p))) |
666e5a40 | 1414 | return true; |
6a46079c | 1415 | |
6a46079c AK |
1416 | /* |
1417 | * This check implies we don't kill processes if their pages | |
1418 | * are in the swap cache early. Those are always late kills. | |
1419 | */ | |
7af446a8 | 1420 | if (!page_mapped(hpage)) |
666e5a40 | 1421 | return true; |
1668bfd5 | 1422 | |
52089b14 | 1423 | if (PageKsm(p)) { |
96f96763 | 1424 | pr_err("%#lx: can't handle KSM pages.\n", pfn); |
666e5a40 | 1425 | return false; |
52089b14 | 1426 | } |
6a46079c AK |
1427 | |
1428 | if (PageSwapCache(p)) { | |
96f96763 | 1429 | pr_err("%#lx: keeping poisoned page in swap cache\n", pfn); |
6a46079c AK |
1430 | ttu |= TTU_IGNORE_HWPOISON; |
1431 | } | |
1432 | ||
1433 | /* | |
1434 | * Propagate the dirty bit from PTEs to struct page first, because we | |
1435 | * need this to decide if we should kill or just drop the page. | |
db0480b3 WF |
1436 | * XXX: the dirty test could be racy: set_page_dirty() may not always |
1437 | * be called inside page lock (it's recommended but not enforced). | |
6a46079c | 1438 | */ |
7af446a8 | 1439 | mapping = page_mapping(hpage); |
6751ed65 | 1440 | if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping && |
f56753ac | 1441 | mapping_can_writeback(mapping)) { |
7af446a8 NH |
1442 | if (page_mkclean(hpage)) { |
1443 | SetPageDirty(hpage); | |
6a46079c AK |
1444 | } else { |
1445 | kill = 0; | |
1446 | ttu |= TTU_IGNORE_HWPOISON; | |
96f96763 | 1447 | pr_info("%#lx: corrupted page was clean: dropped without side effects\n", |
6a46079c AK |
1448 | pfn); |
1449 | } | |
1450 | } | |
1451 | ||
1452 | /* | |
1453 | * First collect all the processes that have the page | |
1454 | * mapped in dirty form. This has to be done before try_to_unmap, | |
1455 | * because ttu takes the rmap data structures down. | |
1456 | * | |
1457 | * Error handling: We ignore errors here because | |
1458 | * there's nothing that can be done. | |
1459 | */ | |
1460 | if (kill) | |
415c64c1 | 1461 | collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED); |
6a46079c | 1462 | |
357670f7 ML |
1463 | if (PageHuge(hpage) && !PageAnon(hpage)) { |
1464 | /* | |
1465 | * For hugetlb pages in shared mappings, try_to_unmap | |
1466 | * could potentially call huge_pmd_unshare. Because of | |
1467 | * this, take semaphore in write mode here and set | |
1468 | * TTU_RMAP_LOCKED to indicate we have taken the lock | |
1469 | * at this higher level. | |
1470 | */ | |
1471 | mapping = hugetlb_page_mapping_lock_write(hpage); | |
1472 | if (mapping) { | |
9030fb0b | 1473 | try_to_unmap(folio, ttu|TTU_RMAP_LOCKED); |
357670f7 ML |
1474 | i_mmap_unlock_write(mapping); |
1475 | } else | |
96f96763 | 1476 | pr_info("%#lx: could not lock mapping for mapped huge page\n", pfn); |
c0d0381a | 1477 | } else { |
9030fb0b | 1478 | try_to_unmap(folio, ttu); |
c0d0381a | 1479 | } |
1fb08ac6 YS |
1480 | |
1481 | unmap_success = !page_mapped(hpage); | |
666e5a40 | 1482 | if (!unmap_success) |
96f96763 | 1483 | pr_err("%#lx: failed to unmap page (mapcount=%d)\n", |
1170532b | 1484 | pfn, page_mapcount(hpage)); |
a6d30ddd | 1485 | |
286c469a NH |
1486 | /* |
1487 | * try_to_unmap() might put mlocked page in lru cache, so call | |
1488 | * shake_page() again to ensure that it's flushed. | |
1489 | */ | |
1490 | if (mlocked) | |
d0505e9f | 1491 | shake_page(hpage); |
286c469a | 1492 | |
6a46079c AK |
1493 | /* |
1494 | * Now that the dirty bit has been propagated to the | |
1495 | * struct page and all unmaps done we can decide if | |
1496 | * killing is needed or not. Only kill when the page | |
6751ed65 TL |
1497 | * was dirty or the process is not restartable, |
1498 | * otherwise the tokill list is merely | |
6a46079c AK |
1499 | * freed. When there was a problem unmapping earlier |
1500 | * use a more force-full uncatchable kill to prevent | |
1501 | * any accesses to the poisoned memory. | |
1502 | */ | |
415c64c1 | 1503 | forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL); |
ae1139ec | 1504 | kill_procs(&tokill, forcekill, !unmap_success, pfn, flags); |
1668bfd5 | 1505 | |
666e5a40 | 1506 | return unmap_success; |
6a46079c AK |
1507 | } |
1508 | ||
0348d2eb NH |
1509 | static int identify_page_state(unsigned long pfn, struct page *p, |
1510 | unsigned long page_flags) | |
761ad8d7 NH |
1511 | { |
1512 | struct page_state *ps; | |
0348d2eb NH |
1513 | |
1514 | /* | |
1515 | * The first check uses the current page flags which may not have any | |
1516 | * relevant information. The second check with the saved page flags is | |
1517 | * carried out only if the first check can't determine the page status. | |
1518 | */ | |
1519 | for (ps = error_states;; ps++) | |
1520 | if ((p->flags & ps->mask) == ps->res) | |
1521 | break; | |
1522 | ||
1523 | page_flags |= (p->flags & (1UL << PG_dirty)); | |
1524 | ||
1525 | if (!ps->mask) | |
1526 | for (ps = error_states;; ps++) | |
1527 | if ((page_flags & ps->mask) == ps->res) | |
1528 | break; | |
1529 | return page_action(ps, p, pfn); | |
1530 | } | |
1531 | ||
694bf0b0 OS |
1532 | static int try_to_split_thp_page(struct page *page, const char *msg) |
1533 | { | |
1534 | lock_page(page); | |
4966455d | 1535 | if (unlikely(split_huge_page(page))) { |
694bf0b0 OS |
1536 | unsigned long pfn = page_to_pfn(page); |
1537 | ||
1538 | unlock_page(page); | |
4966455d | 1539 | pr_info("%s: %#lx: thp split failed\n", msg, pfn); |
694bf0b0 OS |
1540 | put_page(page); |
1541 | return -EBUSY; | |
1542 | } | |
1543 | unlock_page(page); | |
1544 | ||
1545 | return 0; | |
1546 | } | |
1547 | ||
00cc790e SR |
1548 | static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn, |
1549 | struct address_space *mapping, pgoff_t index, int flags) | |
1550 | { | |
1551 | struct to_kill *tk; | |
1552 | unsigned long size = 0; | |
1553 | ||
1554 | list_for_each_entry(tk, to_kill, nd) | |
1555 | if (tk->size_shift) | |
1556 | size = max(size, 1UL << tk->size_shift); | |
1557 | ||
1558 | if (size) { | |
1559 | /* | |
1560 | * Unmap the largest mapping to avoid breaking up device-dax | |
1561 | * mappings which are constant size. The actual size of the | |
1562 | * mapping being torn down is communicated in siginfo, see | |
1563 | * kill_proc() | |
1564 | */ | |
1565 | loff_t start = (index << PAGE_SHIFT) & ~(size - 1); | |
1566 | ||
1567 | unmap_mapping_range(mapping, start, size, 0); | |
1568 | } | |
1569 | ||
1570 | kill_procs(to_kill, flags & MF_MUST_KILL, false, pfn, flags); | |
1571 | } | |
1572 | ||
1573 | static int mf_generic_kill_procs(unsigned long long pfn, int flags, | |
1574 | struct dev_pagemap *pgmap) | |
1575 | { | |
1576 | struct page *page = pfn_to_page(pfn); | |
1577 | LIST_HEAD(to_kill); | |
1578 | dax_entry_t cookie; | |
1579 | int rc = 0; | |
1580 | ||
1581 | /* | |
1582 | * Pages instantiated by device-dax (not filesystem-dax) | |
1583 | * may be compound pages. | |
1584 | */ | |
1585 | page = compound_head(page); | |
1586 | ||
1587 | /* | |
1588 | * Prevent the inode from being freed while we are interrogating | |
1589 | * the address_space, typically this would be handled by | |
1590 | * lock_page(), but dax pages do not use the page lock. This | |
1591 | * also prevents changes to the mapping of this pfn until | |
1592 | * poison signaling is complete. | |
1593 | */ | |
1594 | cookie = dax_lock_page(page); | |
1595 | if (!cookie) | |
1596 | return -EBUSY; | |
1597 | ||
1598 | if (hwpoison_filter(page)) { | |
1599 | rc = -EOPNOTSUPP; | |
1600 | goto unlock; | |
1601 | } | |
1602 | ||
1603 | switch (pgmap->type) { | |
1604 | case MEMORY_DEVICE_PRIVATE: | |
1605 | case MEMORY_DEVICE_COHERENT: | |
1606 | /* | |
1607 | * TODO: Handle device pages which may need coordination | |
1608 | * with device-side memory. | |
1609 | */ | |
1610 | rc = -ENXIO; | |
1611 | goto unlock; | |
1612 | default: | |
1613 | break; | |
1614 | } | |
1615 | ||
1616 | /* | |
1617 | * Use this flag as an indication that the dax page has been | |
1618 | * remapped UC to prevent speculative consumption of poison. | |
1619 | */ | |
1620 | SetPageHWPoison(page); | |
1621 | ||
1622 | /* | |
1623 | * Unlike System-RAM there is no possibility to swap in a | |
1624 | * different physical page at a given virtual address, so all | |
1625 | * userspace consumption of ZONE_DEVICE memory necessitates | |
1626 | * SIGBUS (i.e. MF_MUST_KILL) | |
1627 | */ | |
1628 | flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1629 | collect_procs(page, &to_kill, true); | |
1630 | ||
1631 | unmap_and_kill(&to_kill, pfn, page->mapping, page->index, flags); | |
1632 | unlock: | |
1633 | dax_unlock_page(page, cookie); | |
1634 | return rc; | |
1635 | } | |
1636 | ||
c36e2024 SR |
1637 | #ifdef CONFIG_FS_DAX |
1638 | /** | |
1639 | * mf_dax_kill_procs - Collect and kill processes who are using this file range | |
1640 | * @mapping: address_space of the file in use | |
1641 | * @index: start pgoff of the range within the file | |
1642 | * @count: length of the range, in unit of PAGE_SIZE | |
1643 | * @mf_flags: memory failure flags | |
1644 | */ | |
1645 | int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index, | |
1646 | unsigned long count, int mf_flags) | |
1647 | { | |
1648 | LIST_HEAD(to_kill); | |
1649 | dax_entry_t cookie; | |
1650 | struct page *page; | |
1651 | size_t end = index + count; | |
1652 | ||
1653 | mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL; | |
1654 | ||
1655 | for (; index < end; index++) { | |
1656 | page = NULL; | |
1657 | cookie = dax_lock_mapping_entry(mapping, index, &page); | |
1658 | if (!cookie) | |
1659 | return -EBUSY; | |
1660 | if (!page) | |
1661 | goto unlock; | |
1662 | ||
1663 | SetPageHWPoison(page); | |
1664 | ||
1665 | collect_procs_fsdax(page, mapping, index, &to_kill); | |
1666 | unmap_and_kill(&to_kill, page_to_pfn(page), mapping, | |
1667 | index, mf_flags); | |
1668 | unlock: | |
1669 | dax_unlock_mapping_entry(mapping, index, cookie); | |
1670 | } | |
1671 | return 0; | |
1672 | } | |
1673 | EXPORT_SYMBOL_GPL(mf_dax_kill_procs); | |
1674 | #endif /* CONFIG_FS_DAX */ | |
1675 | ||
161df60e NH |
1676 | #ifdef CONFIG_HUGETLB_PAGE |
1677 | /* | |
1678 | * Struct raw_hwp_page represents information about "raw error page", | |
1679 | * constructing singly linked list originated from ->private field of | |
1680 | * SUBPAGE_INDEX_HWPOISON-th tail page. | |
1681 | */ | |
1682 | struct raw_hwp_page { | |
1683 | struct llist_node node; | |
1684 | struct page *page; | |
1685 | }; | |
1686 | ||
1687 | static inline struct llist_head *raw_hwp_list_head(struct page *hpage) | |
1688 | { | |
1689 | return (struct llist_head *)&page_private(hpage + SUBPAGE_INDEX_HWPOISON); | |
1690 | } | |
1691 | ||
ac5fcde0 | 1692 | static unsigned long __free_raw_hwp_pages(struct page *hpage, bool move_flag) |
161df60e NH |
1693 | { |
1694 | struct llist_head *head; | |
1695 | struct llist_node *t, *tnode; | |
ac5fcde0 | 1696 | unsigned long count = 0; |
161df60e NH |
1697 | |
1698 | head = raw_hwp_list_head(hpage); | |
1699 | llist_for_each_safe(tnode, t, head->first) { | |
1700 | struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node); | |
1701 | ||
ac5fcde0 NH |
1702 | if (move_flag) |
1703 | SetPageHWPoison(p->page); | |
161df60e | 1704 | kfree(p); |
ac5fcde0 | 1705 | count++; |
161df60e NH |
1706 | } |
1707 | llist_del_all(head); | |
ac5fcde0 | 1708 | return count; |
161df60e NH |
1709 | } |
1710 | ||
1711 | static int hugetlb_set_page_hwpoison(struct page *hpage, struct page *page) | |
1712 | { | |
1713 | struct llist_head *head; | |
1714 | struct raw_hwp_page *raw_hwp; | |
1715 | struct llist_node *t, *tnode; | |
1716 | int ret = TestSetPageHWPoison(hpage) ? -EHWPOISON : 0; | |
1717 | ||
1718 | /* | |
1719 | * Once the hwpoison hugepage has lost reliable raw error info, | |
1720 | * there is little meaning to keep additional error info precisely, | |
1721 | * so skip to add additional raw error info. | |
1722 | */ | |
1723 | if (HPageRawHwpUnreliable(hpage)) | |
1724 | return -EHWPOISON; | |
1725 | head = raw_hwp_list_head(hpage); | |
1726 | llist_for_each_safe(tnode, t, head->first) { | |
1727 | struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node); | |
1728 | ||
1729 | if (p->page == page) | |
1730 | return -EHWPOISON; | |
1731 | } | |
1732 | ||
1733 | raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC); | |
1734 | if (raw_hwp) { | |
1735 | raw_hwp->page = page; | |
1736 | llist_add(&raw_hwp->node, head); | |
1737 | /* the first error event will be counted in action_result(). */ | |
1738 | if (ret) | |
1739 | num_poisoned_pages_inc(); | |
1740 | } else { | |
1741 | /* | |
1742 | * Failed to save raw error info. We no longer trace all | |
1743 | * hwpoisoned subpages, and we need refuse to free/dissolve | |
1744 | * this hwpoisoned hugepage. | |
1745 | */ | |
1746 | SetHPageRawHwpUnreliable(hpage); | |
1747 | /* | |
1748 | * Once HPageRawHwpUnreliable is set, raw_hwp_page is not | |
1749 | * used any more, so free it. | |
1750 | */ | |
ac5fcde0 | 1751 | __free_raw_hwp_pages(hpage, false); |
161df60e NH |
1752 | } |
1753 | return ret; | |
1754 | } | |
1755 | ||
ac5fcde0 NH |
1756 | static unsigned long free_raw_hwp_pages(struct page *hpage, bool move_flag) |
1757 | { | |
1758 | /* | |
1759 | * HPageVmemmapOptimized hugepages can't be freed because struct | |
1760 | * pages for tail pages are required but they don't exist. | |
1761 | */ | |
1762 | if (move_flag && HPageVmemmapOptimized(hpage)) | |
1763 | return 0; | |
1764 | ||
1765 | /* | |
1766 | * HPageRawHwpUnreliable hugepages shouldn't be unpoisoned by | |
1767 | * definition. | |
1768 | */ | |
1769 | if (HPageRawHwpUnreliable(hpage)) | |
1770 | return 0; | |
1771 | ||
1772 | return __free_raw_hwp_pages(hpage, move_flag); | |
1773 | } | |
1774 | ||
161df60e NH |
1775 | void hugetlb_clear_page_hwpoison(struct page *hpage) |
1776 | { | |
1777 | if (HPageRawHwpUnreliable(hpage)) | |
1778 | return; | |
1779 | ClearPageHWPoison(hpage); | |
ac5fcde0 | 1780 | free_raw_hwp_pages(hpage, true); |
161df60e NH |
1781 | } |
1782 | ||
405ce051 NH |
1783 | /* |
1784 | * Called from hugetlb code with hugetlb_lock held. | |
1785 | * | |
1786 | * Return values: | |
1787 | * 0 - free hugepage | |
1788 | * 1 - in-use hugepage | |
1789 | * 2 - not a hugepage | |
1790 | * -EBUSY - the hugepage is busy (try to retry) | |
1791 | * -EHWPOISON - the hugepage is already hwpoisoned | |
1792 | */ | |
1793 | int __get_huge_page_for_hwpoison(unsigned long pfn, int flags) | |
1794 | { | |
1795 | struct page *page = pfn_to_page(pfn); | |
1796 | struct page *head = compound_head(page); | |
1797 | int ret = 2; /* fallback to normal page handling */ | |
1798 | bool count_increased = false; | |
1799 | ||
1800 | if (!PageHeadHuge(head)) | |
1801 | goto out; | |
1802 | ||
1803 | if (flags & MF_COUNT_INCREASED) { | |
1804 | ret = 1; | |
1805 | count_increased = true; | |
b283d983 NH |
1806 | } else if (HPageFreed(head)) { |
1807 | ret = 0; | |
1808 | } else if (HPageMigratable(head)) { | |
405ce051 NH |
1809 | ret = get_page_unless_zero(head); |
1810 | if (ret) | |
1811 | count_increased = true; | |
1812 | } else { | |
1813 | ret = -EBUSY; | |
38f6d293 NH |
1814 | if (!(flags & MF_NO_RETRY)) |
1815 | goto out; | |
405ce051 NH |
1816 | } |
1817 | ||
161df60e | 1818 | if (hugetlb_set_page_hwpoison(head, page)) { |
405ce051 NH |
1819 | ret = -EHWPOISON; |
1820 | goto out; | |
1821 | } | |
1822 | ||
1823 | return ret; | |
1824 | out: | |
1825 | if (count_increased) | |
1826 | put_page(head); | |
1827 | return ret; | |
1828 | } | |
1829 | ||
405ce051 NH |
1830 | /* |
1831 | * Taking refcount of hugetlb pages needs extra care about race conditions | |
1832 | * with basic operations like hugepage allocation/free/demotion. | |
1833 | * So some of prechecks for hwpoison (pinning, and testing/setting | |
1834 | * PageHWPoison) should be done in single hugetlb_lock range. | |
1835 | */ | |
1836 | static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) | |
0348d2eb | 1837 | { |
761ad8d7 | 1838 | int res; |
405ce051 NH |
1839 | struct page *p = pfn_to_page(pfn); |
1840 | struct page *head; | |
761ad8d7 NH |
1841 | unsigned long page_flags; |
1842 | ||
405ce051 NH |
1843 | *hugetlb = 1; |
1844 | retry: | |
1845 | res = get_huge_page_for_hwpoison(pfn, flags); | |
1846 | if (res == 2) { /* fallback to normal page handling */ | |
1847 | *hugetlb = 0; | |
1848 | return 0; | |
1849 | } else if (res == -EHWPOISON) { | |
96f96763 | 1850 | pr_err("%#lx: already hardware poisoned\n", pfn); |
405ce051 NH |
1851 | if (flags & MF_ACTION_REQUIRED) { |
1852 | head = compound_head(p); | |
a3f5d80e | 1853 | res = kill_accessing_process(current, page_to_pfn(head), flags); |
405ce051 NH |
1854 | } |
1855 | return res; | |
1856 | } else if (res == -EBUSY) { | |
38f6d293 NH |
1857 | if (!(flags & MF_NO_RETRY)) { |
1858 | flags |= MF_NO_RETRY; | |
405ce051 NH |
1859 | goto retry; |
1860 | } | |
1861 | action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); | |
a3f5d80e | 1862 | return res; |
761ad8d7 NH |
1863 | } |
1864 | ||
405ce051 NH |
1865 | head = compound_head(p); |
1866 | lock_page(head); | |
1867 | ||
1868 | if (hwpoison_filter(p)) { | |
161df60e | 1869 | hugetlb_clear_page_hwpoison(head); |
405ce051 NH |
1870 | res = -EOPNOTSUPP; |
1871 | goto out; | |
1872 | } | |
1873 | ||
405ce051 NH |
1874 | /* |
1875 | * Handling free hugepage. The possible race with hugepage allocation | |
1876 | * or demotion can be prevented by PageHWPoison flag. | |
1877 | */ | |
1878 | if (res == 0) { | |
1879 | unlock_page(head); | |
ceaf8fbe | 1880 | if (__page_handle_poison(p) >= 0) { |
405ce051 NH |
1881 | page_ref_inc(p); |
1882 | res = MF_RECOVERED; | |
ceaf8fbe NH |
1883 | } else { |
1884 | res = MF_FAILED; | |
761ad8d7 | 1885 | } |
405ce051 NH |
1886 | action_result(pfn, MF_MSG_FREE_HUGE, res); |
1887 | return res == MF_RECOVERED ? 0 : -EBUSY; | |
761ad8d7 NH |
1888 | } |
1889 | ||
761ad8d7 NH |
1890 | page_flags = head->flags; |
1891 | ||
ed8c2f49 | 1892 | if (!hwpoison_user_mappings(p, pfn, flags, head)) { |
761ad8d7 NH |
1893 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1894 | res = -EBUSY; | |
1895 | goto out; | |
1896 | } | |
1897 | ||
ea6d0630 | 1898 | return identify_page_state(pfn, p, page_flags); |
761ad8d7 NH |
1899 | out: |
1900 | unlock_page(head); | |
1901 | return res; | |
1902 | } | |
00cc790e | 1903 | |
405ce051 NH |
1904 | #else |
1905 | static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb) | |
1906 | { | |
1907 | return 0; | |
1908 | } | |
00cc790e | 1909 | |
ac5fcde0 NH |
1910 | static inline unsigned long free_raw_hwp_pages(struct page *hpage, bool flag) |
1911 | { | |
1912 | return 0; | |
1913 | } | |
00cc790e | 1914 | #endif /* CONFIG_HUGETLB_PAGE */ |
761ad8d7 | 1915 | |
6100e34b DW |
1916 | static int memory_failure_dev_pagemap(unsigned long pfn, int flags, |
1917 | struct dev_pagemap *pgmap) | |
1918 | { | |
1919 | struct page *page = pfn_to_page(pfn); | |
00cc790e | 1920 | int rc = -ENXIO; |
6100e34b | 1921 | |
1e8aaedb OS |
1922 | if (flags & MF_COUNT_INCREASED) |
1923 | /* | |
1924 | * Drop the extra refcount in case we come from madvise(). | |
1925 | */ | |
1926 | put_page(page); | |
1927 | ||
34dc45be | 1928 | /* device metadata space is not recoverable */ |
00cc790e | 1929 | if (!pgmap_pfn_valid(pgmap, pfn)) |
34dc45be | 1930 | goto out; |
61e28cf0 | 1931 | |
6100e34b | 1932 | /* |
33a8f7f2 SR |
1933 | * Call driver's implementation to handle the memory failure, otherwise |
1934 | * fall back to generic handler. | |
6100e34b | 1935 | */ |
65d3440e | 1936 | if (pgmap_has_memory_failure(pgmap)) { |
33a8f7f2 | 1937 | rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags); |
6100e34b | 1938 | /* |
33a8f7f2 SR |
1939 | * Fall back to generic handler too if operation is not |
1940 | * supported inside the driver/device/filesystem. | |
6100e34b | 1941 | */ |
33a8f7f2 SR |
1942 | if (rc != -EOPNOTSUPP) |
1943 | goto out; | |
6100e34b DW |
1944 | } |
1945 | ||
00cc790e | 1946 | rc = mf_generic_kill_procs(pfn, flags, pgmap); |
6100e34b DW |
1947 | out: |
1948 | /* drop pgmap ref acquired in caller */ | |
1949 | put_dev_pagemap(pgmap); | |
1950 | action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED); | |
1951 | return rc; | |
1952 | } | |
1953 | ||
91d00547 NH |
1954 | static DEFINE_MUTEX(mf_mutex); |
1955 | ||
cd42f4a3 TL |
1956 | /** |
1957 | * memory_failure - Handle memory failure of a page. | |
1958 | * @pfn: Page Number of the corrupted page | |
cd42f4a3 TL |
1959 | * @flags: fine tune action taken |
1960 | * | |
1961 | * This function is called by the low level machine check code | |
1962 | * of an architecture when it detects hardware memory corruption | |
1963 | * of a page. It tries its best to recover, which includes | |
1964 | * dropping pages, killing processes etc. | |
1965 | * | |
1966 | * The function is primarily of use for corruptions that | |
1967 | * happen outside the current execution context (e.g. when | |
1968 | * detected by a background scrubber) | |
1969 | * | |
1970 | * Must run in process context (e.g. a work queue) with interrupts | |
1971 | * enabled and no spinlocks hold. | |
d1fe111f | 1972 | * |
1973 | * Return: 0 for successfully handled the memory error, | |
9113eaf3 | 1974 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event, |
d1fe111f | 1975 | * < 0(except -EOPNOTSUPP) on failure. |
cd42f4a3 | 1976 | */ |
83b57531 | 1977 | int memory_failure(unsigned long pfn, int flags) |
6a46079c | 1978 | { |
6a46079c | 1979 | struct page *p; |
7af446a8 | 1980 | struct page *hpage; |
6100e34b | 1981 | struct dev_pagemap *pgmap; |
171936dd | 1982 | int res = 0; |
524fca1e | 1983 | unsigned long page_flags; |
a8b2c2ce | 1984 | bool retry = true; |
405ce051 | 1985 | int hugetlb = 0; |
6a46079c AK |
1986 | |
1987 | if (!sysctl_memory_failure_recovery) | |
83b57531 | 1988 | panic("Memory failure on page %lx", pfn); |
6a46079c | 1989 | |
03b122da TL |
1990 | mutex_lock(&mf_mutex); |
1991 | ||
67f22ba7 | 1992 | if (!(flags & MF_SW_SIMULATED)) |
1993 | hw_memory_failure = true; | |
1994 | ||
96c804a6 DH |
1995 | p = pfn_to_online_page(pfn); |
1996 | if (!p) { | |
03b122da TL |
1997 | res = arch_memory_failure(pfn, flags); |
1998 | if (res == 0) | |
1999 | goto unlock_mutex; | |
2000 | ||
96c804a6 DH |
2001 | if (pfn_valid(pfn)) { |
2002 | pgmap = get_dev_pagemap(pfn, NULL); | |
03b122da TL |
2003 | if (pgmap) { |
2004 | res = memory_failure_dev_pagemap(pfn, flags, | |
2005 | pgmap); | |
2006 | goto unlock_mutex; | |
2007 | } | |
96c804a6 | 2008 | } |
96f96763 | 2009 | pr_err("%#lx: memory outside kernel control\n", pfn); |
03b122da TL |
2010 | res = -ENXIO; |
2011 | goto unlock_mutex; | |
6a46079c AK |
2012 | } |
2013 | ||
a8b2c2ce | 2014 | try_again: |
405ce051 NH |
2015 | res = try_memory_failure_hugetlb(pfn, flags, &hugetlb); |
2016 | if (hugetlb) | |
171936dd | 2017 | goto unlock_mutex; |
171936dd | 2018 | |
6a46079c | 2019 | if (TestSetPageHWPoison(p)) { |
96f96763 | 2020 | pr_err("%#lx: already hardware poisoned\n", pfn); |
47af12ba | 2021 | res = -EHWPOISON; |
a3f5d80e NH |
2022 | if (flags & MF_ACTION_REQUIRED) |
2023 | res = kill_accessing_process(current, pfn, flags); | |
f361e246 NH |
2024 | if (flags & MF_COUNT_INCREASED) |
2025 | put_page(p); | |
171936dd | 2026 | goto unlock_mutex; |
6a46079c AK |
2027 | } |
2028 | ||
75ee64b3 | 2029 | hpage = compound_head(p); |
6a46079c AK |
2030 | |
2031 | /* | |
2032 | * We need/can do nothing about count=0 pages. | |
2033 | * 1) it's a free page, and therefore in safe hand: | |
2034 | * prep_new_page() will be the gate keeper. | |
761ad8d7 | 2035 | * 2) it's part of a non-compound high order page. |
6a46079c AK |
2036 | * Implies some kernel user: cannot stop them from |
2037 | * R/W the page; let's pray that the page has been | |
2038 | * used and will be freed some time later. | |
2039 | * In fact it's dangerous to directly bump up page count from 0, | |
1c4c3b99 | 2040 | * that may make page_ref_freeze()/page_ref_unfreeze() mismatch. |
6a46079c | 2041 | */ |
0ed950d1 NH |
2042 | if (!(flags & MF_COUNT_INCREASED)) { |
2043 | res = get_hwpoison_page(p, flags); | |
2044 | if (!res) { | |
2045 | if (is_free_buddy_page(p)) { | |
2046 | if (take_page_off_buddy(p)) { | |
2047 | page_ref_inc(p); | |
2048 | res = MF_RECOVERED; | |
2049 | } else { | |
2050 | /* We lost the race, try again */ | |
2051 | if (retry) { | |
2052 | ClearPageHWPoison(p); | |
0ed950d1 NH |
2053 | retry = false; |
2054 | goto try_again; | |
2055 | } | |
2056 | res = MF_FAILED; | |
a8b2c2ce | 2057 | } |
0ed950d1 NH |
2058 | action_result(pfn, MF_MSG_BUDDY, res); |
2059 | res = res == MF_RECOVERED ? 0 : -EBUSY; | |
2060 | } else { | |
2061 | action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED); | |
2062 | res = -EBUSY; | |
a8b2c2ce | 2063 | } |
0ed950d1 NH |
2064 | goto unlock_mutex; |
2065 | } else if (res < 0) { | |
2066 | action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED); | |
171936dd | 2067 | res = -EBUSY; |
0ed950d1 | 2068 | goto unlock_mutex; |
8d22ba1b | 2069 | } |
6a46079c AK |
2070 | } |
2071 | ||
761ad8d7 | 2072 | if (PageTransHuge(hpage)) { |
eac96c3e YS |
2073 | /* |
2074 | * The flag must be set after the refcount is bumped | |
2075 | * otherwise it may race with THP split. | |
2076 | * And the flag can't be set in get_hwpoison_page() since | |
2077 | * it is called by soft offline too and it is just called | |
2078 | * for !MF_COUNT_INCREASE. So here seems to be the best | |
2079 | * place. | |
2080 | * | |
2081 | * Don't need care about the above error handling paths for | |
2082 | * get_hwpoison_page() since they handle either free page | |
2083 | * or unhandlable page. The refcount is bumped iff the | |
2084 | * page is a valid handlable page. | |
2085 | */ | |
2086 | SetPageHasHWPoisoned(hpage); | |
5d1fd5dc NH |
2087 | if (try_to_split_thp_page(p, "Memory Failure") < 0) { |
2088 | action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED); | |
171936dd TL |
2089 | res = -EBUSY; |
2090 | goto unlock_mutex; | |
5d1fd5dc | 2091 | } |
415c64c1 | 2092 | VM_BUG_ON_PAGE(!page_count(p), p); |
415c64c1 NH |
2093 | } |
2094 | ||
e43c3afb WF |
2095 | /* |
2096 | * We ignore non-LRU pages for good reasons. | |
2097 | * - PG_locked is only well defined for LRU pages and a few others | |
48c935ad | 2098 | * - to avoid races with __SetPageLocked() |
e43c3afb WF |
2099 | * - to avoid races with __SetPageSlab*() (and more non-atomic ops) |
2100 | * The check (unnecessarily) ignores LRU pages being isolated and | |
2101 | * walked by the page reclaim code, however that's not a big loss. | |
2102 | */ | |
d0505e9f | 2103 | shake_page(p); |
e43c3afb | 2104 | |
761ad8d7 | 2105 | lock_page(p); |
847ce401 | 2106 | |
f37d4298 | 2107 | /* |
75ee64b3 ML |
2108 | * We're only intended to deal with the non-Compound page here. |
2109 | * However, the page could have changed compound pages due to | |
2110 | * race window. If this happens, we could try again to hopefully | |
2111 | * handle the page next round. | |
f37d4298 | 2112 | */ |
75ee64b3 ML |
2113 | if (PageCompound(p)) { |
2114 | if (retry) { | |
e240ac52 | 2115 | ClearPageHWPoison(p); |
75ee64b3 ML |
2116 | unlock_page(p); |
2117 | put_page(p); | |
2118 | flags &= ~MF_COUNT_INCREASED; | |
2119 | retry = false; | |
2120 | goto try_again; | |
2121 | } | |
cc637b17 | 2122 | action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED); |
f37d4298 | 2123 | res = -EBUSY; |
171936dd | 2124 | goto unlock_page; |
f37d4298 AK |
2125 | } |
2126 | ||
524fca1e NH |
2127 | /* |
2128 | * We use page flags to determine what action should be taken, but | |
2129 | * the flags can be modified by the error containment action. One | |
2130 | * example is an mlocked page, where PG_mlocked is cleared by | |
2131 | * page_remove_rmap() in try_to_unmap_one(). So to determine page status | |
2132 | * correctly, we save a copy of the page flags at this time. | |
2133 | */ | |
7d9d46ac | 2134 | page_flags = p->flags; |
524fca1e | 2135 | |
7c116f2b | 2136 | if (hwpoison_filter(p)) { |
e240ac52 | 2137 | TestClearPageHWPoison(p); |
761ad8d7 | 2138 | unlock_page(p); |
dd6e2402 | 2139 | put_page(p); |
d1fe111f | 2140 | res = -EOPNOTSUPP; |
171936dd | 2141 | goto unlock_mutex; |
7c116f2b | 2142 | } |
847ce401 | 2143 | |
e8675d29 | 2144 | /* |
2145 | * __munlock_pagevec may clear a writeback page's LRU flag without | |
2146 | * page_lock. We need wait writeback completion for this page or it | |
2147 | * may trigger vfs BUG while evict inode. | |
2148 | */ | |
b04d3eeb | 2149 | if (!PageLRU(p) && !PageWriteback(p)) |
0bc1f8b0 CY |
2150 | goto identify_page_state; |
2151 | ||
6edd6cc6 NH |
2152 | /* |
2153 | * It's very difficult to mess with pages currently under IO | |
2154 | * and in many cases impossible, so we just avoid it here. | |
2155 | */ | |
6a46079c AK |
2156 | wait_on_page_writeback(p); |
2157 | ||
2158 | /* | |
2159 | * Now take care of user space mappings. | |
6ffcd825 | 2160 | * Abort on fail: __filemap_remove_folio() assumes unmapped page. |
6a46079c | 2161 | */ |
ed8c2f49 | 2162 | if (!hwpoison_user_mappings(p, pfn, flags, p)) { |
cc637b17 | 2163 | action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED); |
1668bfd5 | 2164 | res = -EBUSY; |
171936dd | 2165 | goto unlock_page; |
1668bfd5 | 2166 | } |
6a46079c AK |
2167 | |
2168 | /* | |
2169 | * Torn down by someone else? | |
2170 | */ | |
dc2a1cbf | 2171 | if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) { |
cc637b17 | 2172 | action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED); |
d95ea51e | 2173 | res = -EBUSY; |
171936dd | 2174 | goto unlock_page; |
6a46079c AK |
2175 | } |
2176 | ||
0bc1f8b0 | 2177 | identify_page_state: |
0348d2eb | 2178 | res = identify_page_state(pfn, p, page_flags); |
ea6d0630 NH |
2179 | mutex_unlock(&mf_mutex); |
2180 | return res; | |
171936dd | 2181 | unlock_page: |
761ad8d7 | 2182 | unlock_page(p); |
171936dd TL |
2183 | unlock_mutex: |
2184 | mutex_unlock(&mf_mutex); | |
6a46079c AK |
2185 | return res; |
2186 | } | |
cd42f4a3 | 2187 | EXPORT_SYMBOL_GPL(memory_failure); |
847ce401 | 2188 | |
ea8f5fb8 HY |
2189 | #define MEMORY_FAILURE_FIFO_ORDER 4 |
2190 | #define MEMORY_FAILURE_FIFO_SIZE (1 << MEMORY_FAILURE_FIFO_ORDER) | |
2191 | ||
2192 | struct memory_failure_entry { | |
2193 | unsigned long pfn; | |
ea8f5fb8 HY |
2194 | int flags; |
2195 | }; | |
2196 | ||
2197 | struct memory_failure_cpu { | |
2198 | DECLARE_KFIFO(fifo, struct memory_failure_entry, | |
2199 | MEMORY_FAILURE_FIFO_SIZE); | |
2200 | spinlock_t lock; | |
2201 | struct work_struct work; | |
2202 | }; | |
2203 | ||
2204 | static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu); | |
2205 | ||
2206 | /** | |
2207 | * memory_failure_queue - Schedule handling memory failure of a page. | |
2208 | * @pfn: Page Number of the corrupted page | |
ea8f5fb8 HY |
2209 | * @flags: Flags for memory failure handling |
2210 | * | |
2211 | * This function is called by the low level hardware error handler | |
2212 | * when it detects hardware memory corruption of a page. It schedules | |
2213 | * the recovering of error page, including dropping pages, killing | |
2214 | * processes etc. | |
2215 | * | |
2216 | * The function is primarily of use for corruptions that | |
2217 | * happen outside the current execution context (e.g. when | |
2218 | * detected by a background scrubber) | |
2219 | * | |
2220 | * Can run in IRQ context. | |
2221 | */ | |
83b57531 | 2222 | void memory_failure_queue(unsigned long pfn, int flags) |
ea8f5fb8 HY |
2223 | { |
2224 | struct memory_failure_cpu *mf_cpu; | |
2225 | unsigned long proc_flags; | |
2226 | struct memory_failure_entry entry = { | |
2227 | .pfn = pfn, | |
ea8f5fb8 HY |
2228 | .flags = flags, |
2229 | }; | |
2230 | ||
2231 | mf_cpu = &get_cpu_var(memory_failure_cpu); | |
2232 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
498d319b | 2233 | if (kfifo_put(&mf_cpu->fifo, entry)) |
ea8f5fb8 HY |
2234 | schedule_work_on(smp_processor_id(), &mf_cpu->work); |
2235 | else | |
96f96763 | 2236 | pr_err("buffer overflow when queuing memory failure at %#lx\n", |
ea8f5fb8 HY |
2237 | pfn); |
2238 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
2239 | put_cpu_var(memory_failure_cpu); | |
2240 | } | |
2241 | EXPORT_SYMBOL_GPL(memory_failure_queue); | |
2242 | ||
2243 | static void memory_failure_work_func(struct work_struct *work) | |
2244 | { | |
2245 | struct memory_failure_cpu *mf_cpu; | |
2246 | struct memory_failure_entry entry = { 0, }; | |
2247 | unsigned long proc_flags; | |
2248 | int gotten; | |
2249 | ||
06202231 | 2250 | mf_cpu = container_of(work, struct memory_failure_cpu, work); |
ea8f5fb8 HY |
2251 | for (;;) { |
2252 | spin_lock_irqsave(&mf_cpu->lock, proc_flags); | |
2253 | gotten = kfifo_get(&mf_cpu->fifo, &entry); | |
2254 | spin_unlock_irqrestore(&mf_cpu->lock, proc_flags); | |
2255 | if (!gotten) | |
2256 | break; | |
cf870c70 | 2257 | if (entry.flags & MF_SOFT_OFFLINE) |
feec24a6 | 2258 | soft_offline_page(entry.pfn, entry.flags); |
cf870c70 | 2259 | else |
83b57531 | 2260 | memory_failure(entry.pfn, entry.flags); |
ea8f5fb8 HY |
2261 | } |
2262 | } | |
2263 | ||
06202231 JM |
2264 | /* |
2265 | * Process memory_failure work queued on the specified CPU. | |
2266 | * Used to avoid return-to-userspace racing with the memory_failure workqueue. | |
2267 | */ | |
2268 | void memory_failure_queue_kick(int cpu) | |
2269 | { | |
2270 | struct memory_failure_cpu *mf_cpu; | |
2271 | ||
2272 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
2273 | cancel_work_sync(&mf_cpu->work); | |
2274 | memory_failure_work_func(&mf_cpu->work); | |
2275 | } | |
2276 | ||
ea8f5fb8 HY |
2277 | static int __init memory_failure_init(void) |
2278 | { | |
2279 | struct memory_failure_cpu *mf_cpu; | |
2280 | int cpu; | |
2281 | ||
2282 | for_each_possible_cpu(cpu) { | |
2283 | mf_cpu = &per_cpu(memory_failure_cpu, cpu); | |
2284 | spin_lock_init(&mf_cpu->lock); | |
2285 | INIT_KFIFO(mf_cpu->fifo); | |
2286 | INIT_WORK(&mf_cpu->work, memory_failure_work_func); | |
2287 | } | |
2288 | ||
2289 | return 0; | |
2290 | } | |
2291 | core_initcall(memory_failure_init); | |
2292 | ||
96f96763 KW |
2293 | #undef pr_fmt |
2294 | #define pr_fmt(fmt) "" fmt | |
a5f65109 NH |
2295 | #define unpoison_pr_info(fmt, pfn, rs) \ |
2296 | ({ \ | |
2297 | if (__ratelimit(rs)) \ | |
2298 | pr_info(fmt, pfn); \ | |
2299 | }) | |
2300 | ||
847ce401 WF |
2301 | /** |
2302 | * unpoison_memory - Unpoison a previously poisoned page | |
2303 | * @pfn: Page number of the to be unpoisoned page | |
2304 | * | |
2305 | * Software-unpoison a page that has been poisoned by | |
2306 | * memory_failure() earlier. | |
2307 | * | |
2308 | * This is only done on the software-level, so it only works | |
2309 | * for linux injected failures, not real hardware failures | |
2310 | * | |
2311 | * Returns 0 for success, otherwise -errno. | |
2312 | */ | |
2313 | int unpoison_memory(unsigned long pfn) | |
2314 | { | |
2315 | struct page *page; | |
2316 | struct page *p; | |
bf181c58 | 2317 | int ret = -EBUSY; |
c8bd84f7 | 2318 | int freeit = 0; |
ac5fcde0 | 2319 | unsigned long count = 1; |
a5f65109 NH |
2320 | static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL, |
2321 | DEFAULT_RATELIMIT_BURST); | |
847ce401 WF |
2322 | |
2323 | if (!pfn_valid(pfn)) | |
2324 | return -ENXIO; | |
2325 | ||
2326 | p = pfn_to_page(pfn); | |
2327 | page = compound_head(p); | |
2328 | ||
91d00547 NH |
2329 | mutex_lock(&mf_mutex); |
2330 | ||
67f22ba7 | 2331 | if (hw_memory_failure) { |
2332 | unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n", | |
2333 | pfn, &unpoison_rs); | |
2334 | ret = -EOPNOTSUPP; | |
2335 | goto unlock_mutex; | |
2336 | } | |
2337 | ||
847ce401 | 2338 | if (!PageHWPoison(p)) { |
495367c0 | 2339 | unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n", |
a5f65109 | 2340 | pfn, &unpoison_rs); |
91d00547 | 2341 | goto unlock_mutex; |
847ce401 WF |
2342 | } |
2343 | ||
230ac719 | 2344 | if (page_count(page) > 1) { |
495367c0 | 2345 | unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n", |
a5f65109 | 2346 | pfn, &unpoison_rs); |
91d00547 | 2347 | goto unlock_mutex; |
230ac719 NH |
2348 | } |
2349 | ||
2350 | if (page_mapped(page)) { | |
495367c0 | 2351 | unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n", |
a5f65109 | 2352 | pfn, &unpoison_rs); |
91d00547 | 2353 | goto unlock_mutex; |
230ac719 NH |
2354 | } |
2355 | ||
2356 | if (page_mapping(page)) { | |
495367c0 | 2357 | unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n", |
a5f65109 | 2358 | pfn, &unpoison_rs); |
91d00547 | 2359 | goto unlock_mutex; |
0cea3fdc WL |
2360 | } |
2361 | ||
bf181c58 | 2362 | if (PageSlab(page) || PageTable(page)) |
91d00547 | 2363 | goto unlock_mutex; |
847ce401 | 2364 | |
bf181c58 NH |
2365 | ret = get_hwpoison_page(p, MF_UNPOISON); |
2366 | if (!ret) { | |
ac5fcde0 NH |
2367 | if (PageHuge(p)) { |
2368 | count = free_raw_hwp_pages(page, false); | |
2369 | if (count == 0) { | |
2370 | ret = -EBUSY; | |
2371 | goto unlock_mutex; | |
2372 | } | |
2373 | } | |
c8bd84f7 | 2374 | ret = TestClearPageHWPoison(page) ? 0 : -EBUSY; |
bf181c58 NH |
2375 | } else if (ret < 0) { |
2376 | if (ret == -EHWPOISON) { | |
c8bd84f7 | 2377 | ret = put_page_back_buddy(p) ? 0 : -EBUSY; |
bf181c58 NH |
2378 | } else |
2379 | unpoison_pr_info("Unpoison: failed to grab page %#lx\n", | |
2380 | pfn, &unpoison_rs); | |
2381 | } else { | |
ac5fcde0 NH |
2382 | if (PageHuge(p)) { |
2383 | count = free_raw_hwp_pages(page, false); | |
2384 | if (count == 0) { | |
2385 | ret = -EBUSY; | |
2386 | goto unlock_mutex; | |
2387 | } | |
2388 | } | |
c8bd84f7 | 2389 | freeit = !!TestClearPageHWPoison(p); |
847ce401 | 2390 | |
dd6e2402 | 2391 | put_page(page); |
bf181c58 NH |
2392 | if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1)) { |
2393 | put_page(page); | |
2394 | ret = 0; | |
2395 | } | |
2396 | } | |
847ce401 | 2397 | |
91d00547 NH |
2398 | unlock_mutex: |
2399 | mutex_unlock(&mf_mutex); | |
c8bd84f7 | 2400 | if (!ret || freeit) { |
ac5fcde0 | 2401 | num_poisoned_pages_sub(count); |
c8bd84f7 | 2402 | unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n", |
2403 | page_to_pfn(p), &unpoison_rs); | |
2404 | } | |
91d00547 | 2405 | return ret; |
847ce401 WF |
2406 | } |
2407 | EXPORT_SYMBOL(unpoison_memory); | |
facb6011 | 2408 | |
6b9a217e | 2409 | static bool isolate_page(struct page *page, struct list_head *pagelist) |
d950b958 | 2410 | { |
6b9a217e OS |
2411 | bool isolated = false; |
2412 | bool lru = PageLRU(page); | |
d950b958 | 2413 | |
6b9a217e | 2414 | if (PageHuge(page)) { |
7ce82f4c | 2415 | isolated = !isolate_hugetlb(page, pagelist); |
6b9a217e OS |
2416 | } else { |
2417 | if (lru) | |
2418 | isolated = !isolate_lru_page(page); | |
2419 | else | |
2420 | isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE); | |
2421 | ||
2422 | if (isolated) | |
2423 | list_add(&page->lru, pagelist); | |
0ebff32c | 2424 | } |
d950b958 | 2425 | |
6b9a217e OS |
2426 | if (isolated && lru) |
2427 | inc_node_page_state(page, NR_ISOLATED_ANON + | |
2428 | page_is_file_lru(page)); | |
2429 | ||
03613808 | 2430 | /* |
6b9a217e OS |
2431 | * If we succeed to isolate the page, we grabbed another refcount on |
2432 | * the page, so we can safely drop the one we got from get_any_pages(). | |
2433 | * If we failed to isolate the page, it means that we cannot go further | |
2434 | * and we will return an error, so drop the reference we got from | |
2435 | * get_any_pages() as well. | |
03613808 | 2436 | */ |
6b9a217e OS |
2437 | put_page(page); |
2438 | return isolated; | |
d950b958 NH |
2439 | } |
2440 | ||
6b9a217e OS |
2441 | /* |
2442 | * __soft_offline_page handles hugetlb-pages and non-hugetlb pages. | |
2443 | * If the page is a non-dirty unmapped page-cache page, it simply invalidates. | |
2444 | * If the page is mapped, it migrates the contents over. | |
2445 | */ | |
2446 | static int __soft_offline_page(struct page *page) | |
af8fae7c | 2447 | { |
d6c75dc2 | 2448 | long ret = 0; |
af8fae7c | 2449 | unsigned long pfn = page_to_pfn(page); |
6b9a217e OS |
2450 | struct page *hpage = compound_head(page); |
2451 | char const *msg_page[] = {"page", "hugepage"}; | |
2452 | bool huge = PageHuge(page); | |
2453 | LIST_HEAD(pagelist); | |
54608759 JK |
2454 | struct migration_target_control mtc = { |
2455 | .nid = NUMA_NO_NODE, | |
2456 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
2457 | }; | |
facb6011 | 2458 | |
0ebff32c | 2459 | lock_page(page); |
6b9a217e OS |
2460 | if (!PageHuge(page)) |
2461 | wait_on_page_writeback(page); | |
af8fae7c NH |
2462 | if (PageHWPoison(page)) { |
2463 | unlock_page(page); | |
dd6e2402 | 2464 | put_page(page); |
af8fae7c | 2465 | pr_info("soft offline: %#lx page already poisoned\n", pfn); |
5a2ffca3 | 2466 | return 0; |
af8fae7c | 2467 | } |
6b9a217e | 2468 | |
593396b8 | 2469 | if (!PageHuge(page) && PageLRU(page) && !PageSwapCache(page)) |
6b9a217e OS |
2470 | /* |
2471 | * Try to invalidate first. This should work for | |
2472 | * non dirty unmapped page cache pages. | |
2473 | */ | |
2474 | ret = invalidate_inode_page(page); | |
facb6011 | 2475 | unlock_page(page); |
6b9a217e | 2476 | |
6b9a217e | 2477 | if (ret) { |
fb46e735 | 2478 | pr_info("soft_offline: %#lx: invalidated\n", pfn); |
6b9a217e | 2479 | page_handle_poison(page, false, true); |
af8fae7c | 2480 | return 0; |
facb6011 AK |
2481 | } |
2482 | ||
6b9a217e | 2483 | if (isolate_page(hpage, &pagelist)) { |
54608759 | 2484 | ret = migrate_pages(&pagelist, alloc_migration_target, NULL, |
5ac95884 | 2485 | (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL); |
79f5f8fa | 2486 | if (!ret) { |
6b9a217e OS |
2487 | bool release = !huge; |
2488 | ||
2489 | if (!page_handle_poison(page, huge, release)) | |
2490 | ret = -EBUSY; | |
79f5f8fa | 2491 | } else { |
85fbe5d1 YX |
2492 | if (!list_empty(&pagelist)) |
2493 | putback_movable_pages(&pagelist); | |
59c82b70 | 2494 | |
d6c75dc2 | 2495 | pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n", |
23efd080 | 2496 | pfn, msg_page[huge], ret, &page->flags); |
facb6011 | 2497 | if (ret > 0) |
3f4b815a | 2498 | ret = -EBUSY; |
facb6011 AK |
2499 | } |
2500 | } else { | |
23efd080 MWO |
2501 | pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n", |
2502 | pfn, msg_page[huge], page_count(page), &page->flags); | |
6b9a217e | 2503 | ret = -EBUSY; |
facb6011 | 2504 | } |
facb6011 AK |
2505 | return ret; |
2506 | } | |
86e05773 | 2507 | |
6b9a217e | 2508 | static int soft_offline_in_use_page(struct page *page) |
acc14dc4 | 2509 | { |
acc14dc4 NH |
2510 | struct page *hpage = compound_head(page); |
2511 | ||
694bf0b0 OS |
2512 | if (!PageHuge(page) && PageTransHuge(hpage)) |
2513 | if (try_to_split_thp_page(page, "soft offline") < 0) | |
acc14dc4 | 2514 | return -EBUSY; |
6b9a217e | 2515 | return __soft_offline_page(page); |
acc14dc4 NH |
2516 | } |
2517 | ||
d4ae9916 | 2518 | static int soft_offline_free_page(struct page *page) |
acc14dc4 | 2519 | { |
6b9a217e | 2520 | int rc = 0; |
acc14dc4 | 2521 | |
6b9a217e OS |
2522 | if (!page_handle_poison(page, true, false)) |
2523 | rc = -EBUSY; | |
06be6ff3 | 2524 | |
d4ae9916 | 2525 | return rc; |
acc14dc4 NH |
2526 | } |
2527 | ||
dad4e5b3 DW |
2528 | static void put_ref_page(struct page *page) |
2529 | { | |
2530 | if (page) | |
2531 | put_page(page); | |
2532 | } | |
2533 | ||
86e05773 WL |
2534 | /** |
2535 | * soft_offline_page - Soft offline a page. | |
feec24a6 | 2536 | * @pfn: pfn to soft-offline |
86e05773 WL |
2537 | * @flags: flags. Same as memory_failure(). |
2538 | * | |
9113eaf3 | 2539 | * Returns 0 on success |
2540 | * -EOPNOTSUPP for hwpoison_filter() filtered the error event | |
2541 | * < 0 otherwise negated errno. | |
86e05773 WL |
2542 | * |
2543 | * Soft offline a page, by migration or invalidation, | |
2544 | * without killing anything. This is for the case when | |
2545 | * a page is not corrupted yet (so it's still valid to access), | |
2546 | * but has had a number of corrected errors and is better taken | |
2547 | * out. | |
2548 | * | |
2549 | * The actual policy on when to do that is maintained by | |
2550 | * user space. | |
2551 | * | |
2552 | * This should never impact any application or cause data loss, | |
2553 | * however it might take some time. | |
2554 | * | |
2555 | * This is not a 100% solution for all memory, but tries to be | |
2556 | * ``good enough'' for the majority of memory. | |
2557 | */ | |
feec24a6 | 2558 | int soft_offline_page(unsigned long pfn, int flags) |
86e05773 WL |
2559 | { |
2560 | int ret; | |
b94e0282 | 2561 | bool try_again = true; |
dad4e5b3 DW |
2562 | struct page *page, *ref_page = NULL; |
2563 | ||
2564 | WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED)); | |
86e05773 | 2565 | |
feec24a6 NH |
2566 | if (!pfn_valid(pfn)) |
2567 | return -ENXIO; | |
dad4e5b3 DW |
2568 | if (flags & MF_COUNT_INCREASED) |
2569 | ref_page = pfn_to_page(pfn); | |
2570 | ||
feec24a6 NH |
2571 | /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */ |
2572 | page = pfn_to_online_page(pfn); | |
dad4e5b3 DW |
2573 | if (!page) { |
2574 | put_ref_page(ref_page); | |
86a66810 | 2575 | return -EIO; |
dad4e5b3 | 2576 | } |
86a66810 | 2577 | |
91d00547 NH |
2578 | mutex_lock(&mf_mutex); |
2579 | ||
86e05773 | 2580 | if (PageHWPoison(page)) { |
8295d535 | 2581 | pr_info("%s: %#lx page already poisoned\n", __func__, pfn); |
dad4e5b3 | 2582 | put_ref_page(ref_page); |
91d00547 | 2583 | mutex_unlock(&mf_mutex); |
5a2ffca3 | 2584 | return 0; |
86e05773 | 2585 | } |
86e05773 | 2586 | |
b94e0282 | 2587 | retry: |
bfc8c901 | 2588 | get_online_mems(); |
bf6445bc | 2589 | ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE); |
bfc8c901 | 2590 | put_online_mems(); |
4e41a30c | 2591 | |
9113eaf3 | 2592 | if (hwpoison_filter(page)) { |
2593 | if (ret > 0) | |
2594 | put_page(page); | |
2595 | else | |
2596 | put_ref_page(ref_page); | |
2597 | ||
2598 | mutex_unlock(&mf_mutex); | |
2599 | return -EOPNOTSUPP; | |
2600 | } | |
2601 | ||
8295d535 | 2602 | if (ret > 0) { |
6b9a217e | 2603 | ret = soft_offline_in_use_page(page); |
8295d535 | 2604 | } else if (ret == 0) { |
b94e0282 OS |
2605 | if (soft_offline_free_page(page) && try_again) { |
2606 | try_again = false; | |
2a57d83c | 2607 | flags &= ~MF_COUNT_INCREASED; |
b94e0282 OS |
2608 | goto retry; |
2609 | } | |
8295d535 | 2610 | } |
4e41a30c | 2611 | |
91d00547 NH |
2612 | mutex_unlock(&mf_mutex); |
2613 | ||
86e05773 WL |
2614 | return ret; |
2615 | } | |
60f272f6 | 2616 | |
2617 | void clear_hwpoisoned_pages(struct page *memmap, int nr_pages) | |
2618 | { | |
2619 | int i; | |
2620 | ||
2621 | /* | |
2622 | * A further optimization is to have per section refcounted | |
2623 | * num_poisoned_pages. But that would need more space per memmap, so | |
2624 | * for now just do a quick global check to speed up this routine in the | |
2625 | * absence of bad pages. | |
2626 | */ | |
2627 | if (atomic_long_read(&num_poisoned_pages) == 0) | |
2628 | return; | |
2629 | ||
2630 | for (i = 0; i < nr_pages; i++) { | |
2631 | if (PageHWPoison(&memmap[i])) { | |
2632 | num_poisoned_pages_dec(); | |
2633 | ClearPageHWPoison(&memmap[i]); | |
2634 | } | |
2635 | } | |
2636 | } |