xref: /linux/arch/x86/kernel/cpu/mce/dev-mcelog.c (revision 4be5e8648b0c287aefc6ac3f3a0b12c696054f43)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * /dev/mcelog driver
4  *
5  * K8 parts Copyright 2002,2003 Andi Kleen, SuSE Labs.
6  * Rest from unknown author(s).
7  * 2004 Andi Kleen. Rewrote most of it.
8  * Copyright 2008 Intel Corporation
9  * Author: Andi Kleen
10  */
11 
12 #include <linux/miscdevice.h>
13 #include <linux/slab.h>
14 #include <linux/kmod.h>
15 #include <linux/poll.h>
16 
17 #include "internal.h"
18 
19 static BLOCKING_NOTIFIER_HEAD(mce_injector_chain);
20 
21 static DEFINE_MUTEX(mce_chrdev_read_mutex);
22 
23 static char mce_helper[128];
24 static char *mce_helper_argv[2] = { mce_helper, NULL };
25 
26 /*
27  * Lockless MCE logging infrastructure.
28  * This avoids deadlocks on printk locks without having to break locks. Also
29  * separate MCEs from kernel messages to avoid bogus bug reports.
30  */
31 
32 static struct mce_log_buffer *mcelog;
33 
34 static DECLARE_WAIT_QUEUE_HEAD(mce_chrdev_wait);
35 
36 static int dev_mce_log(struct notifier_block *nb, unsigned long val,
37 				void *data)
38 {
39 	struct mce *mce = (struct mce *)data;
40 	unsigned int entry;
41 
42 	mutex_lock(&mce_chrdev_read_mutex);
43 
44 	entry = mcelog->next;
45 
46 	/*
47 	 * When the buffer fills up discard new entries. Assume that the
48 	 * earlier errors are the more interesting ones:
49 	 */
50 	if (entry >= mcelog->len) {
51 		set_bit(MCE_OVERFLOW, (unsigned long *)&mcelog->flags);
52 		goto unlock;
53 	}
54 
55 	mcelog->next = entry + 1;
56 
57 	memcpy(mcelog->entry + entry, mce, sizeof(struct mce));
58 	mcelog->entry[entry].finished = 1;
59 
60 	/* wake processes polling /dev/mcelog */
61 	wake_up_interruptible(&mce_chrdev_wait);
62 
63 unlock:
64 	mutex_unlock(&mce_chrdev_read_mutex);
65 
66 	return NOTIFY_OK;
67 }
68 
69 static struct notifier_block dev_mcelog_nb = {
70 	.notifier_call	= dev_mce_log,
71 	.priority	= MCE_PRIO_MCELOG,
72 };
73 
74 static void mce_do_trigger(struct work_struct *work)
75 {
76 	call_usermodehelper(mce_helper, mce_helper_argv, NULL, UMH_NO_WAIT);
77 }
78 
79 static DECLARE_WORK(mce_trigger_work, mce_do_trigger);
80 
81 
82 void mce_work_trigger(void)
83 {
84 	if (mce_helper[0])
85 		schedule_work(&mce_trigger_work);
86 }
87 
88 static ssize_t
89 show_trigger(struct device *s, struct device_attribute *attr, char *buf)
90 {
91 	strcpy(buf, mce_helper);
92 	strcat(buf, "\n");
93 	return strlen(mce_helper) + 1;
94 }
95 
96 static ssize_t set_trigger(struct device *s, struct device_attribute *attr,
97 				const char *buf, size_t siz)
98 {
99 	char *p;
100 
101 	strncpy(mce_helper, buf, sizeof(mce_helper));
102 	mce_helper[sizeof(mce_helper)-1] = 0;
103 	p = strchr(mce_helper, '\n');
104 
105 	if (p)
106 		*p = 0;
107 
108 	return strlen(mce_helper) + !!p;
109 }
110 
111 DEVICE_ATTR(trigger, 0644, show_trigger, set_trigger);
112 
113 /*
114  * mce_chrdev: Character device /dev/mcelog to read and clear the MCE log.
115  */
116 
117 static DEFINE_SPINLOCK(mce_chrdev_state_lock);
118 static int mce_chrdev_open_count;	/* #times opened */
119 static int mce_chrdev_open_exclu;	/* already open exclusive? */
120 
121 static int mce_chrdev_open(struct inode *inode, struct file *file)
122 {
123 	spin_lock(&mce_chrdev_state_lock);
124 
125 	if (mce_chrdev_open_exclu ||
126 	    (mce_chrdev_open_count && (file->f_flags & O_EXCL))) {
127 		spin_unlock(&mce_chrdev_state_lock);
128 
129 		return -EBUSY;
130 	}
131 
132 	if (file->f_flags & O_EXCL)
133 		mce_chrdev_open_exclu = 1;
134 	mce_chrdev_open_count++;
135 
136 	spin_unlock(&mce_chrdev_state_lock);
137 
138 	return nonseekable_open(inode, file);
139 }
140 
141 static int mce_chrdev_release(struct inode *inode, struct file *file)
142 {
143 	spin_lock(&mce_chrdev_state_lock);
144 
145 	mce_chrdev_open_count--;
146 	mce_chrdev_open_exclu = 0;
147 
148 	spin_unlock(&mce_chrdev_state_lock);
149 
150 	return 0;
151 }
152 
153 static int mce_apei_read_done;
154 
155 /* Collect MCE record of previous boot in persistent storage via APEI ERST. */
156 static int __mce_read_apei(char __user **ubuf, size_t usize)
157 {
158 	int rc;
159 	u64 record_id;
160 	struct mce m;
161 
162 	if (usize < sizeof(struct mce))
163 		return -EINVAL;
164 
165 	rc = apei_read_mce(&m, &record_id);
166 	/* Error or no more MCE record */
167 	if (rc <= 0) {
168 		mce_apei_read_done = 1;
169 		/*
170 		 * When ERST is disabled, mce_chrdev_read() should return
171 		 * "no record" instead of "no device."
172 		 */
173 		if (rc == -ENODEV)
174 			return 0;
175 		return rc;
176 	}
177 	rc = -EFAULT;
178 	if (copy_to_user(*ubuf, &m, sizeof(struct mce)))
179 		return rc;
180 	/*
181 	 * In fact, we should have cleared the record after that has
182 	 * been flushed to the disk or sent to network in
183 	 * /sbin/mcelog, but we have no interface to support that now,
184 	 * so just clear it to avoid duplication.
185 	 */
186 	rc = apei_clear_mce(record_id);
187 	if (rc) {
188 		mce_apei_read_done = 1;
189 		return rc;
190 	}
191 	*ubuf += sizeof(struct mce);
192 
193 	return 0;
194 }
195 
196 static ssize_t mce_chrdev_read(struct file *filp, char __user *ubuf,
197 				size_t usize, loff_t *off)
198 {
199 	char __user *buf = ubuf;
200 	unsigned next;
201 	int i, err;
202 
203 	mutex_lock(&mce_chrdev_read_mutex);
204 
205 	if (!mce_apei_read_done) {
206 		err = __mce_read_apei(&buf, usize);
207 		if (err || buf != ubuf)
208 			goto out;
209 	}
210 
211 	/* Only supports full reads right now */
212 	err = -EINVAL;
213 	if (*off != 0 || usize < mcelog->len * sizeof(struct mce))
214 		goto out;
215 
216 	next = mcelog->next;
217 	err = 0;
218 
219 	for (i = 0; i < next; i++) {
220 		struct mce *m = &mcelog->entry[i];
221 
222 		err |= copy_to_user(buf, m, sizeof(*m));
223 		buf += sizeof(*m);
224 	}
225 
226 	memset(mcelog->entry, 0, next * sizeof(struct mce));
227 	mcelog->next = 0;
228 
229 	if (err)
230 		err = -EFAULT;
231 
232 out:
233 	mutex_unlock(&mce_chrdev_read_mutex);
234 
235 	return err ? err : buf - ubuf;
236 }
237 
238 static __poll_t mce_chrdev_poll(struct file *file, poll_table *wait)
239 {
240 	poll_wait(file, &mce_chrdev_wait, wait);
241 	if (READ_ONCE(mcelog->next))
242 		return EPOLLIN | EPOLLRDNORM;
243 	if (!mce_apei_read_done && apei_check_mce())
244 		return EPOLLIN | EPOLLRDNORM;
245 	return 0;
246 }
247 
248 static long mce_chrdev_ioctl(struct file *f, unsigned int cmd,
249 				unsigned long arg)
250 {
251 	int __user *p = (int __user *)arg;
252 
253 	if (!capable(CAP_SYS_ADMIN))
254 		return -EPERM;
255 
256 	switch (cmd) {
257 	case MCE_GET_RECORD_LEN:
258 		return put_user(sizeof(struct mce), p);
259 	case MCE_GET_LOG_LEN:
260 		return put_user(mcelog->len, p);
261 	case MCE_GETCLEAR_FLAGS: {
262 		unsigned flags;
263 
264 		do {
265 			flags = mcelog->flags;
266 		} while (cmpxchg(&mcelog->flags, flags, 0) != flags);
267 
268 		return put_user(flags, p);
269 	}
270 	default:
271 		return -ENOTTY;
272 	}
273 }
274 
275 void mce_register_injector_chain(struct notifier_block *nb)
276 {
277 	blocking_notifier_chain_register(&mce_injector_chain, nb);
278 }
279 EXPORT_SYMBOL_GPL(mce_register_injector_chain);
280 
281 void mce_unregister_injector_chain(struct notifier_block *nb)
282 {
283 	blocking_notifier_chain_unregister(&mce_injector_chain, nb);
284 }
285 EXPORT_SYMBOL_GPL(mce_unregister_injector_chain);
286 
287 static ssize_t mce_chrdev_write(struct file *filp, const char __user *ubuf,
288 				size_t usize, loff_t *off)
289 {
290 	struct mce m;
291 
292 	if (!capable(CAP_SYS_ADMIN))
293 		return -EPERM;
294 	/*
295 	 * There are some cases where real MSR reads could slip
296 	 * through.
297 	 */
298 	if (!boot_cpu_has(X86_FEATURE_MCE) || !boot_cpu_has(X86_FEATURE_MCA))
299 		return -EIO;
300 
301 	if ((unsigned long)usize > sizeof(struct mce))
302 		usize = sizeof(struct mce);
303 	if (copy_from_user(&m, ubuf, usize))
304 		return -EFAULT;
305 
306 	if (m.extcpu >= num_possible_cpus() || !cpu_online(m.extcpu))
307 		return -EINVAL;
308 
309 	/*
310 	 * Need to give user space some time to set everything up,
311 	 * so do it a jiffie or two later everywhere.
312 	 */
313 	schedule_timeout(2);
314 
315 	blocking_notifier_call_chain(&mce_injector_chain, 0, &m);
316 
317 	return usize;
318 }
319 
320 static const struct file_operations mce_chrdev_ops = {
321 	.open			= mce_chrdev_open,
322 	.release		= mce_chrdev_release,
323 	.read			= mce_chrdev_read,
324 	.write			= mce_chrdev_write,
325 	.poll			= mce_chrdev_poll,
326 	.unlocked_ioctl		= mce_chrdev_ioctl,
327 	.llseek			= no_llseek,
328 };
329 
330 static struct miscdevice mce_chrdev_device = {
331 	MISC_MCELOG_MINOR,
332 	"mcelog",
333 	&mce_chrdev_ops,
334 };
335 
336 static __init int dev_mcelog_init_device(void)
337 {
338 	int mce_log_len;
339 	int err;
340 
341 	mce_log_len = max(MCE_LOG_MIN_LEN, num_online_cpus());
342 	mcelog = kzalloc(sizeof(*mcelog) + mce_log_len * sizeof(struct mce), GFP_KERNEL);
343 	if (!mcelog)
344 		return -ENOMEM;
345 
346 	strncpy(mcelog->signature, MCE_LOG_SIGNATURE, sizeof(mcelog->signature));
347 	mcelog->len = mce_log_len;
348 	mcelog->recordlen = sizeof(struct mce);
349 
350 	/* register character device /dev/mcelog */
351 	err = misc_register(&mce_chrdev_device);
352 	if (err) {
353 		if (err == -EBUSY)
354 			/* Xen dom0 might have registered the device already. */
355 			pr_info("Unable to init device /dev/mcelog, already registered");
356 		else
357 			pr_err("Unable to init device /dev/mcelog (rc: %d)\n", err);
358 
359 		kfree(mcelog);
360 		return err;
361 	}
362 
363 	mce_register_decode_chain(&dev_mcelog_nb);
364 	return 0;
365 }
366 device_initcall_sync(dev_mcelog_init_device);
367