aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--Documentation/sysctl/vm.txt41
-rw-r--r--fs/proc/meminfo.c9
-rw-r--r--include/linux/mm.h7
-rw-r--r--include/linux/rmap.h1
-rw-r--r--kernel/sysctl.c25
-rw-r--r--mm/Kconfig10
-rw-r--r--mm/Makefile1
-rw-r--r--mm/filemap.c4
-rw-r--r--mm/memory-failure.c832
-rw-r--r--mm/rmap.c7
10 files changed, 934 insertions, 3 deletions
diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
index c4de6359d440..faf62740aa2c 100644
--- a/Documentation/sysctl/vm.txt
+++ b/Documentation/sysctl/vm.txt
@@ -32,6 +32,8 @@ Currently, these files are in /proc/sys/vm:
32- legacy_va_layout 32- legacy_va_layout
33- lowmem_reserve_ratio 33- lowmem_reserve_ratio
34- max_map_count 34- max_map_count
35- memory_failure_early_kill
36- memory_failure_recovery
35- min_free_kbytes 37- min_free_kbytes
36- min_slab_ratio 38- min_slab_ratio
37- min_unmapped_ratio 39- min_unmapped_ratio
@@ -53,7 +55,6 @@ Currently, these files are in /proc/sys/vm:
53- vfs_cache_pressure 55- vfs_cache_pressure
54- zone_reclaim_mode 56- zone_reclaim_mode
55 57
56
57============================================================== 58==============================================================
58 59
59block_dump 60block_dump
@@ -275,6 +276,44 @@ e.g., up to one or two maps per allocation.
275 276
276The default value is 65536. 277The default value is 65536.
277 278
279=============================================================
280
281memory_failure_early_kill:
282
283Control how to kill processes when uncorrected memory error (typically
284a 2bit error in a memory module) is detected in the background by hardware
285that cannot be handled by the kernel. In some cases (like the page
286still having a valid copy on disk) the kernel will handle the failure
287transparently without affecting any applications. But if there is
288no other uptodate copy of the data it will kill to prevent any data
289corruptions from propagating.
290
2911: Kill all processes that have the corrupted and not reloadable page mapped
292as soon as the corruption is detected. Note this is not supported
293for a few types of pages, like kernel internally allocated data or
294the swap cache, but works for the majority of user pages.
295
2960: Only unmap the corrupted page from all processes and only kill a process
297who tries to access it.
298
299The kill is done using a catchable SIGBUS with BUS_MCEERR_AO, so processes can
300handle this if they want to.
301
302This is only active on architectures/platforms with advanced machine
303check handling and depends on the hardware capabilities.
304
305Applications can override this setting individually with the PR_MCE_KILL prctl
306
307==============================================================
308
309memory_failure_recovery
310
311Enable memory failure recovery (when supported by the platform)
312
3131: Attempt recovery.
314
3150: Always panic on a memory failure.
316
278============================================================== 317==============================================================
279 318
280min_free_kbytes: 319min_free_kbytes:
diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c
index d5c410d47fae..78faedcb0a8d 100644
--- a/fs/proc/meminfo.c
+++ b/fs/proc/meminfo.c
@@ -95,7 +95,11 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
95 "Committed_AS: %8lu kB\n" 95 "Committed_AS: %8lu kB\n"
96 "VmallocTotal: %8lu kB\n" 96 "VmallocTotal: %8lu kB\n"
97 "VmallocUsed: %8lu kB\n" 97 "VmallocUsed: %8lu kB\n"
98 "VmallocChunk: %8lu kB\n", 98 "VmallocChunk: %8lu kB\n"
99#ifdef CONFIG_MEMORY_FAILURE
100 "HardwareCorrupted: %8lu kB\n"
101#endif
102 ,
99 K(i.totalram), 103 K(i.totalram),
100 K(i.freeram), 104 K(i.freeram),
101 K(i.bufferram), 105 K(i.bufferram),
@@ -140,6 +144,9 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
140 (unsigned long)VMALLOC_TOTAL >> 10, 144 (unsigned long)VMALLOC_TOTAL >> 10,
141 vmi.used >> 10, 145 vmi.used >> 10,
142 vmi.largest_chunk >> 10 146 vmi.largest_chunk >> 10
147#ifdef CONFIG_MEMORY_FAILURE
148 ,atomic_long_read(&mce_bad_pages) << (PAGE_SHIFT - 10)
149#endif
143 ); 150 );
144 151
145 hugetlb_report_meminfo(m); 152 hugetlb_report_meminfo(m);
diff --git a/include/linux/mm.h b/include/linux/mm.h
index a16018f7d61c..1ffca03f34b7 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1309,5 +1309,12 @@ void vmemmap_populate_print_last(void);
1309extern int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim, 1309extern int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim,
1310 size_t size); 1310 size_t size);
1311extern void refund_locked_memory(struct mm_struct *mm, size_t size); 1311extern void refund_locked_memory(struct mm_struct *mm, size_t size);
1312
1313extern void memory_failure(unsigned long pfn, int trapno);
1314extern int __memory_failure(unsigned long pfn, int trapno, int ref);
1315extern int sysctl_memory_failure_early_kill;
1316extern int sysctl_memory_failure_recovery;
1317extern atomic_long_t mce_bad_pages;
1318
1312#endif /* __KERNEL__ */ 1319#endif /* __KERNEL__ */
1313#endif /* _LINUX_MM_H */ 1320#endif /* _LINUX_MM_H */
diff --git a/include/linux/rmap.h b/include/linux/rmap.h
index ce989f1fc2ed..3c1004e50747 100644
--- a/include/linux/rmap.h
+++ b/include/linux/rmap.h
@@ -129,6 +129,7 @@ int try_to_munlock(struct page *);
129 */ 129 */
130struct anon_vma *page_lock_anon_vma(struct page *page); 130struct anon_vma *page_lock_anon_vma(struct page *page);
131void page_unlock_anon_vma(struct anon_vma *anon_vma); 131void page_unlock_anon_vma(struct anon_vma *anon_vma);
132int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma);
132 133
133#else /* !CONFIG_MMU */ 134#else /* !CONFIG_MMU */
134 135
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index 6bb59f707402..eacae77ac9fc 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -1372,6 +1372,31 @@ static struct ctl_table vm_table[] = {
1372 .mode = 0644, 1372 .mode = 0644,
1373 .proc_handler = &scan_unevictable_handler, 1373 .proc_handler = &scan_unevictable_handler,
1374 }, 1374 },
1375#ifdef CONFIG_MEMORY_FAILURE
1376 {
1377 .ctl_name = CTL_UNNUMBERED,
1378 .procname = "memory_failure_early_kill",
1379 .data = &sysctl_memory_failure_early_kill,
1380 .maxlen = sizeof(sysctl_memory_failure_early_kill),
1381 .mode = 0644,
1382 .proc_handler = &proc_dointvec_minmax,
1383 .strategy = &sysctl_intvec,
1384 .extra1 = &zero,
1385 .extra2 = &one,
1386 },
1387 {
1388 .ctl_name = CTL_UNNUMBERED,
1389 .procname = "memory_failure_recovery",
1390 .data = &sysctl_memory_failure_recovery,
1391 .maxlen = sizeof(sysctl_memory_failure_recovery),
1392 .mode = 0644,
1393 .proc_handler = &proc_dointvec_minmax,
1394 .strategy = &sysctl_intvec,
1395 .extra1 = &zero,
1396 .extra2 = &one,
1397 },
1398#endif
1399
1375/* 1400/*
1376 * NOTE: do not add new entries to this table unless you have read 1401 * NOTE: do not add new entries to this table unless you have read
1377 * Documentation/sysctl/ctl_unnumbered.txt 1402 * Documentation/sysctl/ctl_unnumbered.txt
diff --git a/mm/Kconfig b/mm/Kconfig
index 3aa519f52e18..ea2d8b61c631 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -233,6 +233,16 @@ config DEFAULT_MMAP_MIN_ADDR
233 /proc/sys/vm/mmap_min_addr tunable. 233 /proc/sys/vm/mmap_min_addr tunable.
234 234
235 235
236config MEMORY_FAILURE
237 depends on MMU
238 depends on X86_MCE
239 bool "Enable recovery from hardware memory errors"
240 help
241 Enables code to recover from some memory failures on systems
242 with MCA recovery. This allows a system to continue running
243 even when some of its memory has uncorrected errors. This requires
244 special hardware support and typically ECC memory.
245
236config NOMMU_INITIAL_TRIM_EXCESS 246config NOMMU_INITIAL_TRIM_EXCESS
237 int "Turn on mmap() excess space trimming before booting" 247 int "Turn on mmap() excess space trimming before booting"
238 depends on !MMU 248 depends on !MMU
diff --git a/mm/Makefile b/mm/Makefile
index ea4b18bd3960..dc2551e7d006 100644
--- a/mm/Makefile
+++ b/mm/Makefile
@@ -40,5 +40,6 @@ obj-$(CONFIG_SMP) += allocpercpu.o
40endif 40endif
41obj-$(CONFIG_QUICKLIST) += quicklist.o 41obj-$(CONFIG_QUICKLIST) += quicklist.o
42obj-$(CONFIG_CGROUP_MEM_RES_CTLR) += memcontrol.o page_cgroup.o 42obj-$(CONFIG_CGROUP_MEM_RES_CTLR) += memcontrol.o page_cgroup.o
43obj-$(CONFIG_MEMORY_FAILURE) += memory-failure.o
43obj-$(CONFIG_DEBUG_KMEMLEAK) += kmemleak.o 44obj-$(CONFIG_DEBUG_KMEMLEAK) += kmemleak.o
44obj-$(CONFIG_DEBUG_KMEMLEAK_TEST) += kmemleak-test.o 45obj-$(CONFIG_DEBUG_KMEMLEAK_TEST) += kmemleak-test.o
diff --git a/mm/filemap.c b/mm/filemap.c
index dd51c68e2b86..75575c392167 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -104,6 +104,10 @@
104 * 104 *
105 * ->task->proc_lock 105 * ->task->proc_lock
106 * ->dcache_lock (proc_pid_lookup) 106 * ->dcache_lock (proc_pid_lookup)
107 *
108 * (code doesn't rely on that order, so you could switch it around)
109 * ->tasklist_lock (memory_failure, collect_procs_ao)
110 * ->i_mmap_lock
107 */ 111 */
108 112
109/* 113/*
diff --git a/mm/memory-failure.c b/mm/memory-failure.c
new file mode 100644
index 000000000000..729d4b15b645
--- /dev/null
+++ b/mm/memory-failure.c
@@ -0,0 +1,832 @@
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 2bit ECC memory or cache
11 * failure.
12 *
13 * Handles page cache pages in various states. The tricky part
14 * here is that we can access any page asynchronous to other VM
15 * users, because memory failures could happen anytime and anywhere,
16 * possibly violating some of their assumptions. This is why this code
17 * has to be extremely careful. Generally it tries to use normal locking
18 * rules, as in get the standard locks, even if that means the
19 * error handling takes potentially a long time.
20 *
21 * The operation to map back from RMAP chains to processes has to walk
22 * the complete process list and has non linear complexity with the number
23 * mappings. In short it can be quite slow. But since memory corruptions
24 * are rare we hope to get away with this.
25 */
26
27/*
28 * Notebook:
29 * - hugetlb needs more code
30 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
31 * - pass bad pages to kdump next kernel
32 */
33#define DEBUG 1 /* remove me in 2.6.34 */
34#include <linux/kernel.h>
35#include <linux/mm.h>
36#include <linux/page-flags.h>
37#include <linux/sched.h>
38#include <linux/rmap.h>
39#include <linux/pagemap.h>
40#include <linux/swap.h>
41#include <linux/backing-dev.h>
42#include "internal.h"
43
44int sysctl_memory_failure_early_kill __read_mostly = 0;
45
46int sysctl_memory_failure_recovery __read_mostly = 1;
47
48atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
49
50/*
51 * Send all the processes who have the page mapped an ``action optional''
52 * signal.
53 */
54static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
55 unsigned long pfn)
56{
57 struct siginfo si;
58 int ret;
59
60 printk(KERN_ERR
61 "MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
62 pfn, t->comm, t->pid);
63 si.si_signo = SIGBUS;
64 si.si_errno = 0;
65 si.si_code = BUS_MCEERR_AO;
66 si.si_addr = (void *)addr;
67#ifdef __ARCH_SI_TRAPNO
68 si.si_trapno = trapno;
69#endif
70 si.si_addr_lsb = PAGE_SHIFT;
71 /*
72 * Don't use force here, it's convenient if the signal
73 * can be temporarily blocked.
74 * This could cause a loop when the user sets SIGBUS
75 * to SIG_IGN, but hopefully noone will do that?
76 */
77 ret = send_sig_info(SIGBUS, &si, t); /* synchronous? */
78 if (ret < 0)
79 printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
80 t->comm, t->pid, ret);
81 return ret;
82}
83
84/*
85 * Kill all processes that have a poisoned page mapped and then isolate
86 * the page.
87 *
88 * General strategy:
89 * Find all processes having the page mapped and kill them.
90 * But we keep a page reference around so that the page is not
91 * actually freed yet.
92 * Then stash the page away
93 *
94 * There's no convenient way to get back to mapped processes
95 * from the VMAs. So do a brute-force search over all
96 * running processes.
97 *
98 * Remember that machine checks are not common (or rather
99 * if they are common you have other problems), so this shouldn't
100 * be a performance issue.
101 *
102 * Also there are some races possible while we get from the
103 * error detection to actually handle it.
104 */
105
106struct to_kill {
107 struct list_head nd;
108 struct task_struct *tsk;
109 unsigned long addr;
110 unsigned addr_valid:1;
111};
112
113/*
114 * Failure handling: if we can't find or can't kill a process there's
115 * not much we can do. We just print a message and ignore otherwise.
116 */
117
118/*
119 * Schedule a process for later kill.
120 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
121 * TBD would GFP_NOIO be enough?
122 */
123static void add_to_kill(struct task_struct *tsk, struct page *p,
124 struct vm_area_struct *vma,
125 struct list_head *to_kill,
126 struct to_kill **tkc)
127{
128 struct to_kill *tk;
129
130 if (*tkc) {
131 tk = *tkc;
132 *tkc = NULL;
133 } else {
134 tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
135 if (!tk) {
136 printk(KERN_ERR
137 "MCE: Out of memory while machine check handling\n");
138 return;
139 }
140 }
141 tk->addr = page_address_in_vma(p, vma);
142 tk->addr_valid = 1;
143
144 /*
145 * In theory we don't have to kill when the page was
146 * munmaped. But it could be also a mremap. Since that's
147 * likely very rare kill anyways just out of paranoia, but use
148 * a SIGKILL because the error is not contained anymore.
149 */
150 if (tk->addr == -EFAULT) {
151 pr_debug("MCE: Unable to find user space address %lx in %s\n",
152 page_to_pfn(p), tsk->comm);
153 tk->addr_valid = 0;
154 }
155 get_task_struct(tsk);
156 tk->tsk = tsk;
157 list_add_tail(&tk->nd, to_kill);
158}
159
160/*
161 * Kill the processes that have been collected earlier.
162 *
163 * Only do anything when DOIT is set, otherwise just free the list
164 * (this is used for clean pages which do not need killing)
165 * Also when FAIL is set do a force kill because something went
166 * wrong earlier.
167 */
168static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
169 int fail, unsigned long pfn)
170{
171 struct to_kill *tk, *next;
172
173 list_for_each_entry_safe (tk, next, to_kill, nd) {
174 if (doit) {
175 /*
176 * In case something went wrong with munmaping
177 * make sure the process doesn't catch the
178 * signal and then access the memory. Just kill it.
179 * the signal handlers
180 */
181 if (fail || tk->addr_valid == 0) {
182 printk(KERN_ERR
183 "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
184 pfn, tk->tsk->comm, tk->tsk->pid);
185 force_sig(SIGKILL, tk->tsk);
186 }
187
188 /*
189 * In theory the process could have mapped
190 * something else on the address in-between. We could
191 * check for that, but we need to tell the
192 * process anyways.
193 */
194 else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
195 pfn) < 0)
196 printk(KERN_ERR
197 "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
198 pfn, tk->tsk->comm, tk->tsk->pid);
199 }
200 put_task_struct(tk->tsk);
201 kfree(tk);
202 }
203}
204
205static int task_early_kill(struct task_struct *tsk)
206{
207 if (!tsk->mm)
208 return 0;
209 if (tsk->flags & PF_MCE_PROCESS)
210 return !!(tsk->flags & PF_MCE_EARLY);
211 return sysctl_memory_failure_early_kill;
212}
213
214/*
215 * Collect processes when the error hit an anonymous page.
216 */
217static void collect_procs_anon(struct page *page, struct list_head *to_kill,
218 struct to_kill **tkc)
219{
220 struct vm_area_struct *vma;
221 struct task_struct *tsk;
222 struct anon_vma *av;
223
224 read_lock(&tasklist_lock);
225 av = page_lock_anon_vma(page);
226 if (av == NULL) /* Not actually mapped anymore */
227 goto out;
228 for_each_process (tsk) {
229 if (!task_early_kill(tsk))
230 continue;
231 list_for_each_entry (vma, &av->head, anon_vma_node) {
232 if (!page_mapped_in_vma(page, vma))
233 continue;
234 if (vma->vm_mm == tsk->mm)
235 add_to_kill(tsk, page, vma, to_kill, tkc);
236 }
237 }
238 page_unlock_anon_vma(av);
239out:
240 read_unlock(&tasklist_lock);
241}
242
243/*
244 * Collect processes when the error hit a file mapped page.
245 */
246static void collect_procs_file(struct page *page, struct list_head *to_kill,
247 struct to_kill **tkc)
248{
249 struct vm_area_struct *vma;
250 struct task_struct *tsk;
251 struct prio_tree_iter iter;
252 struct address_space *mapping = page->mapping;
253
254 /*
255 * A note on the locking order between the two locks.
256 * We don't rely on this particular order.
257 * If you have some other code that needs a different order
258 * feel free to switch them around. Or add a reverse link
259 * from mm_struct to task_struct, then this could be all
260 * done without taking tasklist_lock and looping over all tasks.
261 */
262
263 read_lock(&tasklist_lock);
264 spin_lock(&mapping->i_mmap_lock);
265 for_each_process(tsk) {
266 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
267
268 if (!task_early_kill(tsk))
269 continue;
270
271 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
272 pgoff) {
273 /*
274 * Send early kill signal to tasks where a vma covers
275 * the page but the corrupted page is not necessarily
276 * mapped it in its pte.
277 * Assume applications who requested early kill want
278 * to be informed of all such data corruptions.
279 */
280 if (vma->vm_mm == tsk->mm)
281 add_to_kill(tsk, page, vma, to_kill, tkc);
282 }
283 }
284 spin_unlock(&mapping->i_mmap_lock);
285 read_unlock(&tasklist_lock);
286}
287
288/*
289 * Collect the processes who have the corrupted page mapped to kill.
290 * This is done in two steps for locking reasons.
291 * First preallocate one tokill structure outside the spin locks,
292 * so that we can kill at least one process reasonably reliable.
293 */
294static void collect_procs(struct page *page, struct list_head *tokill)
295{
296 struct to_kill *tk;
297
298 if (!page->mapping)
299 return;
300
301 tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
302 if (!tk)
303 return;
304 if (PageAnon(page))
305 collect_procs_anon(page, tokill, &tk);
306 else
307 collect_procs_file(page, tokill, &tk);
308 kfree(tk);
309}
310
311/*
312 * Error handlers for various types of pages.
313 */
314
315enum outcome {
316 FAILED, /* Error handling failed */
317 DELAYED, /* Will be handled later */
318 IGNORED, /* Error safely ignored */
319 RECOVERED, /* Successfully recovered */
320};
321
322static const char *action_name[] = {
323 [FAILED] = "Failed",
324 [DELAYED] = "Delayed",
325 [IGNORED] = "Ignored",
326 [RECOVERED] = "Recovered",
327};
328
329/*
330 * Error hit kernel page.
331 * Do nothing, try to be lucky and not touch this instead. For a few cases we
332 * could be more sophisticated.
333 */
334static int me_kernel(struct page *p, unsigned long pfn)
335{
336 return DELAYED;
337}
338
339/*
340 * Already poisoned page.
341 */
342static int me_ignore(struct page *p, unsigned long pfn)
343{
344 return IGNORED;
345}
346
347/*
348 * Page in unknown state. Do nothing.
349 */
350static int me_unknown(struct page *p, unsigned long pfn)
351{
352 printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
353 return FAILED;
354}
355
356/*
357 * Free memory
358 */
359static int me_free(struct page *p, unsigned long pfn)
360{
361 return DELAYED;
362}
363
364/*
365 * Clean (or cleaned) page cache page.
366 */
367static int me_pagecache_clean(struct page *p, unsigned long pfn)
368{
369 int err;
370 int ret = FAILED;
371 struct address_space *mapping;
372
373 if (!isolate_lru_page(p))
374 page_cache_release(p);
375
376 /*
377 * For anonymous pages we're done the only reference left
378 * should be the one m_f() holds.
379 */
380 if (PageAnon(p))
381 return RECOVERED;
382
383 /*
384 * Now truncate the page in the page cache. This is really
385 * more like a "temporary hole punch"
386 * Don't do this for block devices when someone else
387 * has a reference, because it could be file system metadata
388 * and that's not safe to truncate.
389 */
390 mapping = page_mapping(p);
391 if (!mapping) {
392 /*
393 * Page has been teared down in the meanwhile
394 */
395 return FAILED;
396 }
397
398 /*
399 * Truncation is a bit tricky. Enable it per file system for now.
400 *
401 * Open: to take i_mutex or not for this? Right now we don't.
402 */
403 if (mapping->a_ops->error_remove_page) {
404 err = mapping->a_ops->error_remove_page(mapping, p);
405 if (err != 0) {
406 printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
407 pfn, err);
408 } else if (page_has_private(p) &&
409 !try_to_release_page(p, GFP_NOIO)) {
410 pr_debug("MCE %#lx: failed to release buffers\n", pfn);
411 } else {
412 ret = RECOVERED;
413 }
414 } else {
415 /*
416 * If the file system doesn't support it just invalidate
417 * This fails on dirty or anything with private pages
418 */
419 if (invalidate_inode_page(p))
420 ret = RECOVERED;
421 else
422 printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
423 pfn);
424 }
425 return ret;
426}
427
428/*
429 * Dirty cache page page
430 * Issues: when the error hit a hole page the error is not properly
431 * propagated.
432 */
433static int me_pagecache_dirty(struct page *p, unsigned long pfn)
434{
435 struct address_space *mapping = page_mapping(p);
436
437 SetPageError(p);
438 /* TBD: print more information about the file. */
439 if (mapping) {
440 /*
441 * IO error will be reported by write(), fsync(), etc.
442 * who check the mapping.
443 * This way the application knows that something went
444 * wrong with its dirty file data.
445 *
446 * There's one open issue:
447 *
448 * The EIO will be only reported on the next IO
449 * operation and then cleared through the IO map.
450 * Normally Linux has two mechanisms to pass IO error
451 * first through the AS_EIO flag in the address space
452 * and then through the PageError flag in the page.
453 * Since we drop pages on memory failure handling the
454 * only mechanism open to use is through AS_AIO.
455 *
456 * This has the disadvantage that it gets cleared on
457 * the first operation that returns an error, while
458 * the PageError bit is more sticky and only cleared
459 * when the page is reread or dropped. If an
460 * application assumes it will always get error on
461 * fsync, but does other operations on the fd before
462 * and the page is dropped inbetween then the error
463 * will not be properly reported.
464 *
465 * This can already happen even without hwpoisoned
466 * pages: first on metadata IO errors (which only
467 * report through AS_EIO) or when the page is dropped
468 * at the wrong time.
469 *
470 * So right now we assume that the application DTRT on
471 * the first EIO, but we're not worse than other parts
472 * of the kernel.
473 */
474 mapping_set_error(mapping, EIO);
475 }
476
477 return me_pagecache_clean(p, pfn);
478}
479
480/*
481 * Clean and dirty swap cache.
482 *
483 * Dirty swap cache page is tricky to handle. The page could live both in page
484 * cache and swap cache(ie. page is freshly swapped in). So it could be
485 * referenced concurrently by 2 types of PTEs:
486 * normal PTEs and swap PTEs. We try to handle them consistently by calling
487 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
488 * and then
489 * - clear dirty bit to prevent IO
490 * - remove from LRU
491 * - but keep in the swap cache, so that when we return to it on
492 * a later page fault, we know the application is accessing
493 * corrupted data and shall be killed (we installed simple
494 * interception code in do_swap_page to catch it).
495 *
496 * Clean swap cache pages can be directly isolated. A later page fault will
497 * bring in the known good data from disk.
498 */
499static int me_swapcache_dirty(struct page *p, unsigned long pfn)
500{
501 int ret = FAILED;
502
503 ClearPageDirty(p);
504 /* Trigger EIO in shmem: */
505 ClearPageUptodate(p);
506
507 if (!isolate_lru_page(p)) {
508 page_cache_release(p);
509 ret = DELAYED;
510 }
511
512 return ret;
513}
514
515static int me_swapcache_clean(struct page *p, unsigned long pfn)
516{
517 int ret = FAILED;
518
519 if (!isolate_lru_page(p)) {
520 page_cache_release(p);
521 ret = RECOVERED;
522 }
523 delete_from_swap_cache(p);
524 return ret;
525}
526
527/*
528 * Huge pages. Needs work.
529 * Issues:
530 * No rmap support so we cannot find the original mapper. In theory could walk
531 * all MMs and look for the mappings, but that would be non atomic and racy.
532 * Need rmap for hugepages for this. Alternatively we could employ a heuristic,
533 * like just walking the current process and hoping it has it mapped (that
534 * should be usually true for the common "shared database cache" case)
535 * Should handle free huge pages and dequeue them too, but this needs to
536 * handle huge page accounting correctly.
537 */
538static int me_huge_page(struct page *p, unsigned long pfn)
539{
540 return FAILED;
541}
542
543/*
544 * Various page states we can handle.
545 *
546 * A page state is defined by its current page->flags bits.
547 * The table matches them in order and calls the right handler.
548 *
549 * This is quite tricky because we can access page at any time
550 * in its live cycle, so all accesses have to be extremly careful.
551 *
552 * This is not complete. More states could be added.
553 * For any missing state don't attempt recovery.
554 */
555
556#define dirty (1UL << PG_dirty)
557#define sc (1UL << PG_swapcache)
558#define unevict (1UL << PG_unevictable)
559#define mlock (1UL << PG_mlocked)
560#define writeback (1UL << PG_writeback)
561#define lru (1UL << PG_lru)
562#define swapbacked (1UL << PG_swapbacked)
563#define head (1UL << PG_head)
564#define tail (1UL << PG_tail)
565#define compound (1UL << PG_compound)
566#define slab (1UL << PG_slab)
567#define buddy (1UL << PG_buddy)
568#define reserved (1UL << PG_reserved)
569
570static struct page_state {
571 unsigned long mask;
572 unsigned long res;
573 char *msg;
574 int (*action)(struct page *p, unsigned long pfn);
575} error_states[] = {
576 { reserved, reserved, "reserved kernel", me_ignore },
577 { buddy, buddy, "free kernel", me_free },
578
579 /*
580 * Could in theory check if slab page is free or if we can drop
581 * currently unused objects without touching them. But just
582 * treat it as standard kernel for now.
583 */
584 { slab, slab, "kernel slab", me_kernel },
585
586#ifdef CONFIG_PAGEFLAGS_EXTENDED
587 { head, head, "huge", me_huge_page },
588 { tail, tail, "huge", me_huge_page },
589#else
590 { compound, compound, "huge", me_huge_page },
591#endif
592
593 { sc|dirty, sc|dirty, "swapcache", me_swapcache_dirty },
594 { sc|dirty, sc, "swapcache", me_swapcache_clean },
595
596 { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
597 { unevict, unevict, "unevictable LRU", me_pagecache_clean},
598
599#ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
600 { mlock|dirty, mlock|dirty, "mlocked LRU", me_pagecache_dirty },
601 { mlock, mlock, "mlocked LRU", me_pagecache_clean },
602#endif
603
604 { lru|dirty, lru|dirty, "LRU", me_pagecache_dirty },
605 { lru|dirty, lru, "clean LRU", me_pagecache_clean },
606 { swapbacked, swapbacked, "anonymous", me_pagecache_clean },
607
608 /*
609 * Catchall entry: must be at end.
610 */
611 { 0, 0, "unknown page state", me_unknown },
612};
613
614#undef lru
615
616static void action_result(unsigned long pfn, char *msg, int result)
617{
618 struct page *page = NULL;
619 if (pfn_valid(pfn))
620 page = pfn_to_page(pfn);
621
622 printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
623 pfn,
624 page && PageDirty(page) ? "dirty " : "",
625 msg, action_name[result]);
626}
627
628static int page_action(struct page_state *ps, struct page *p,
629 unsigned long pfn, int ref)
630{
631 int result;
632
633 result = ps->action(p, pfn);
634 action_result(pfn, ps->msg, result);
635 if (page_count(p) != 1 + ref)
636 printk(KERN_ERR
637 "MCE %#lx: %s page still referenced by %d users\n",
638 pfn, ps->msg, page_count(p) - 1);
639
640 /* Could do more checks here if page looks ok */
641 /*
642 * Could adjust zone counters here to correct for the missing page.
643 */
644
645 return result == RECOVERED ? 0 : -EBUSY;
646}
647
648#define N_UNMAP_TRIES 5
649
650/*
651 * Do all that is necessary to remove user space mappings. Unmap
652 * the pages and send SIGBUS to the processes if the data was dirty.
653 */
654static void hwpoison_user_mappings(struct page *p, unsigned long pfn,
655 int trapno)
656{
657 enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
658 struct address_space *mapping;
659 LIST_HEAD(tokill);
660 int ret;
661 int i;
662 int kill = 1;
663
664 if (PageReserved(p) || PageCompound(p) || PageSlab(p))
665 return;
666
667 if (!PageLRU(p))
668 lru_add_drain_all();
669
670 /*
671 * This check implies we don't kill processes if their pages
672 * are in the swap cache early. Those are always late kills.
673 */
674 if (!page_mapped(p))
675 return;
676
677 if (PageSwapCache(p)) {
678 printk(KERN_ERR
679 "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
680 ttu |= TTU_IGNORE_HWPOISON;
681 }
682
683 /*
684 * Propagate the dirty bit from PTEs to struct page first, because we
685 * need this to decide if we should kill or just drop the page.
686 */
687 mapping = page_mapping(p);
688 if (!PageDirty(p) && mapping && mapping_cap_writeback_dirty(mapping)) {
689 if (page_mkclean(p)) {
690 SetPageDirty(p);
691 } else {
692 kill = 0;
693 ttu |= TTU_IGNORE_HWPOISON;
694 printk(KERN_INFO
695 "MCE %#lx: corrupted page was clean: dropped without side effects\n",
696 pfn);
697 }
698 }
699
700 /*
701 * First collect all the processes that have the page
702 * mapped in dirty form. This has to be done before try_to_unmap,
703 * because ttu takes the rmap data structures down.
704 *
705 * Error handling: We ignore errors here because
706 * there's nothing that can be done.
707 */
708 if (kill)
709 collect_procs(p, &tokill);
710
711 /*
712 * try_to_unmap can fail temporarily due to races.
713 * Try a few times (RED-PEN better strategy?)
714 */
715 for (i = 0; i < N_UNMAP_TRIES; i++) {
716 ret = try_to_unmap(p, ttu);
717 if (ret == SWAP_SUCCESS)
718 break;
719 pr_debug("MCE %#lx: try_to_unmap retry needed %d\n", pfn, ret);
720 }
721
722 if (ret != SWAP_SUCCESS)
723 printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
724 pfn, page_mapcount(p));
725
726 /*
727 * Now that the dirty bit has been propagated to the
728 * struct page and all unmaps done we can decide if
729 * killing is needed or not. Only kill when the page
730 * was dirty, otherwise the tokill list is merely
731 * freed. When there was a problem unmapping earlier
732 * use a more force-full uncatchable kill to prevent
733 * any accesses to the poisoned memory.
734 */
735 kill_procs_ao(&tokill, !!PageDirty(p), trapno,
736 ret != SWAP_SUCCESS, pfn);
737}
738
739int __memory_failure(unsigned long pfn, int trapno, int ref)
740{
741 struct page_state *ps;
742 struct page *p;
743 int res;
744
745 if (!sysctl_memory_failure_recovery)
746 panic("Memory failure from trap %d on page %lx", trapno, pfn);
747
748 if (!pfn_valid(pfn)) {
749 action_result(pfn, "memory outside kernel control", IGNORED);
750 return -EIO;
751 }
752
753 p = pfn_to_page(pfn);
754 if (TestSetPageHWPoison(p)) {
755 action_result(pfn, "already hardware poisoned", IGNORED);
756 return 0;
757 }
758
759 atomic_long_add(1, &mce_bad_pages);
760
761 /*
762 * We need/can do nothing about count=0 pages.
763 * 1) it's a free page, and therefore in safe hand:
764 * prep_new_page() will be the gate keeper.
765 * 2) it's part of a non-compound high order page.
766 * Implies some kernel user: cannot stop them from
767 * R/W the page; let's pray that the page has been
768 * used and will be freed some time later.
769 * In fact it's dangerous to directly bump up page count from 0,
770 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
771 */
772 if (!get_page_unless_zero(compound_head(p))) {
773 action_result(pfn, "free or high order kernel", IGNORED);
774 return PageBuddy(compound_head(p)) ? 0 : -EBUSY;
775 }
776
777 /*
778 * Lock the page and wait for writeback to finish.
779 * It's very difficult to mess with pages currently under IO
780 * and in many cases impossible, so we just avoid it here.
781 */
782 lock_page_nosync(p);
783 wait_on_page_writeback(p);
784
785 /*
786 * Now take care of user space mappings.
787 */
788 hwpoison_user_mappings(p, pfn, trapno);
789
790 /*
791 * Torn down by someone else?
792 */
793 if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
794 action_result(pfn, "already truncated LRU", IGNORED);
795 res = 0;
796 goto out;
797 }
798
799 res = -EBUSY;
800 for (ps = error_states;; ps++) {
801 if ((p->flags & ps->mask) == ps->res) {
802 res = page_action(ps, p, pfn, ref);
803 break;
804 }
805 }
806out:
807 unlock_page(p);
808 return res;
809}
810EXPORT_SYMBOL_GPL(__memory_failure);
811
812/**
813 * memory_failure - Handle memory failure of a page.
814 * @pfn: Page Number of the corrupted page
815 * @trapno: Trap number reported in the signal to user space.
816 *
817 * This function is called by the low level machine check code
818 * of an architecture when it detects hardware memory corruption
819 * of a page. It tries its best to recover, which includes
820 * dropping pages, killing processes etc.
821 *
822 * The function is primarily of use for corruptions that
823 * happen outside the current execution context (e.g. when
824 * detected by a background scrubber)
825 *
826 * Must run in process context (e.g. a work queue) with interrupts
827 * enabled and no spinlocks hold.
828 */
829void memory_failure(unsigned long pfn, int trapno)
830{
831 __memory_failure(pfn, trapno, 0);
832}
diff --git a/mm/rmap.c b/mm/rmap.c
index 7e72ca19d68b..09c3d0b96116 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -36,6 +36,11 @@
36 * mapping->tree_lock (widely used, in set_page_dirty, 36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock, 37 * in arch-dependent flush_dcache_mmap_lock,
38 * within inode_lock in __sync_single_inode) 38 * within inode_lock in __sync_single_inode)
39 *
40 * (code doesn't rely on that order so it could be switched around)
41 * ->tasklist_lock
42 * anon_vma->lock (memory_failure, collect_procs_anon)
43 * pte map lock
39 */ 44 */
40 45
41#include <linux/mm.h> 46#include <linux/mm.h>
@@ -311,7 +316,7 @@ pte_t *page_check_address(struct page *page, struct mm_struct *mm,
311 * if the page is not mapped into the page tables of this VMA. Only 316 * if the page is not mapped into the page tables of this VMA. Only
312 * valid for normal file or anonymous VMAs. 317 * valid for normal file or anonymous VMAs.
313 */ 318 */
314static int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) 319int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
315{ 320{
316 unsigned long address; 321 unsigned long address;
317 pte_t *pte; 322 pte_t *pte;