path: root/mm/memory-failure.c (plain)
blob: d6524dce43b26690aa098c3afeefb23541ee59e1

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