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
19static BLOCKING_NOTIFIER_HEAD(mce_injector_chain);
20
21static DEFINE_MUTEX(mce_chrdev_read_mutex);
22
23static char mce_helper[128];
24static 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
32static struct mce_log_buffer *mcelog;
33
34static DECLARE_WAIT_QUEUE_HEAD(mce_chrdev_wait);
35
36static 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	if (mce->kflags & MCE_HANDLED_CEC)
43		return NOTIFY_DONE;
44
45	mutex_lock(&mce_chrdev_read_mutex);
46
47	entry = mcelog->next;
48
49	/*
50	 * When the buffer fills up discard new entries. Assume that the
51	 * earlier errors are the more interesting ones:
52	 */
53	if (entry >= mcelog->len) {
54		set_bit(MCE_OVERFLOW, (unsigned long *)&mcelog->flags);
55		goto unlock;
56	}
57
58	mcelog->next = entry + 1;
59
60	memcpy(mcelog->entry + entry, mce, sizeof(struct mce));
61	mcelog->entry[entry].finished = 1;
62	mcelog->entry[entry].kflags = 0;
63
64	/* wake processes polling /dev/mcelog */
65	wake_up_interruptible(&mce_chrdev_wait);
66
67unlock:
68	mutex_unlock(&mce_chrdev_read_mutex);
69
70	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
71		mce->kflags |= MCE_HANDLED_MCELOG;
72
73	return NOTIFY_OK;
74}
75
76static struct notifier_block dev_mcelog_nb = {
77	.notifier_call	= dev_mce_log,
78	.priority	= MCE_PRIO_MCELOG,
79};
80
81static void mce_do_trigger(struct work_struct *work)
82{
83	call_usermodehelper(mce_helper, mce_helper_argv, NULL, UMH_NO_WAIT);
84}
85
86static DECLARE_WORK(mce_trigger_work, mce_do_trigger);
87
88
89void mce_work_trigger(void)
90{
91	if (mce_helper[0])
92		schedule_work(&mce_trigger_work);
93}
94
95static ssize_t
96show_trigger(struct device *s, struct device_attribute *attr, char *buf)
97{
98	strcpy(buf, mce_helper);
99	strcat(buf, "\n");
100	return strlen(mce_helper) + 1;
101}
102
103static ssize_t set_trigger(struct device *s, struct device_attribute *attr,
104				const char *buf, size_t siz)
105{
106	char *p;
107
108	strscpy(mce_helper, buf, sizeof(mce_helper));
109	p = strchr(mce_helper, '\n');
110
111	if (p)
112		*p = 0;
113
114	return strlen(mce_helper) + !!p;
115}
116
117DEVICE_ATTR(trigger, 0644, show_trigger, set_trigger);
118
119/*
120 * mce_chrdev: Character device /dev/mcelog to read and clear the MCE log.
121 */
122
123static DEFINE_SPINLOCK(mce_chrdev_state_lock);
124static int mce_chrdev_open_count;	/* #times opened */
125static int mce_chrdev_open_exclu;	/* already open exclusive? */
126
127static int mce_chrdev_open(struct inode *inode, struct file *file)
128{
129	spin_lock(&mce_chrdev_state_lock);
130
131	if (mce_chrdev_open_exclu ||
132	    (mce_chrdev_open_count && (file->f_flags & O_EXCL))) {
133		spin_unlock(&mce_chrdev_state_lock);
134
135		return -EBUSY;
136	}
137
138	if (file->f_flags & O_EXCL)
139		mce_chrdev_open_exclu = 1;
140	mce_chrdev_open_count++;
141
142	spin_unlock(&mce_chrdev_state_lock);
143
144	return nonseekable_open(inode, file);
145}
146
147static int mce_chrdev_release(struct inode *inode, struct file *file)
148{
149	spin_lock(&mce_chrdev_state_lock);
150
151	mce_chrdev_open_count--;
152	mce_chrdev_open_exclu = 0;
153
154	spin_unlock(&mce_chrdev_state_lock);
155
156	return 0;
157}
158
159static int mce_apei_read_done;
160
161/* Collect MCE record of previous boot in persistent storage via APEI ERST. */
162static int __mce_read_apei(char __user **ubuf, size_t usize)
163{
164	int rc;
165	u64 record_id;
166	struct mce m;
167
168	if (usize < sizeof(struct mce))
169		return -EINVAL;
170
171	rc = apei_read_mce(&m, &record_id);
172	/* Error or no more MCE record */
173	if (rc <= 0) {
174		mce_apei_read_done = 1;
175		/*
176		 * When ERST is disabled, mce_chrdev_read() should return
177		 * "no record" instead of "no device."
178		 */
179		if (rc == -ENODEV)
180			return 0;
181		return rc;
182	}
183	rc = -EFAULT;
184	if (copy_to_user(*ubuf, &m, sizeof(struct mce)))
185		return rc;
186	/*
187	 * In fact, we should have cleared the record after that has
188	 * been flushed to the disk or sent to network in
189	 * /sbin/mcelog, but we have no interface to support that now,
190	 * so just clear it to avoid duplication.
191	 */
192	rc = apei_clear_mce(record_id);
193	if (rc) {
194		mce_apei_read_done = 1;
195		return rc;
196	}
197	*ubuf += sizeof(struct mce);
198
199	return 0;
200}
201
202static ssize_t mce_chrdev_read(struct file *filp, char __user *ubuf,
203				size_t usize, loff_t *off)
204{
205	char __user *buf = ubuf;
206	unsigned next;
207	int i, err;
208
209	mutex_lock(&mce_chrdev_read_mutex);
210
211	if (!mce_apei_read_done) {
212		err = __mce_read_apei(&buf, usize);
213		if (err || buf != ubuf)
214			goto out;
215	}
216
217	/* Only supports full reads right now */
218	err = -EINVAL;
219	if (*off != 0 || usize < mcelog->len * sizeof(struct mce))
220		goto out;
221
222	next = mcelog->next;
223	err = 0;
224
225	for (i = 0; i < next; i++) {
226		struct mce *m = &mcelog->entry[i];
227
228		err |= copy_to_user(buf, m, sizeof(*m));
229		buf += sizeof(*m);
230	}
231
232	memset(mcelog->entry, 0, next * sizeof(struct mce));
233	mcelog->next = 0;
234
235	if (err)
236		err = -EFAULT;
237
238out:
239	mutex_unlock(&mce_chrdev_read_mutex);
240
241	return err ? err : buf - ubuf;
242}
243
244static __poll_t mce_chrdev_poll(struct file *file, poll_table *wait)
245{
246	poll_wait(file, &mce_chrdev_wait, wait);
247	if (READ_ONCE(mcelog->next))
248		return EPOLLIN | EPOLLRDNORM;
249	if (!mce_apei_read_done && apei_check_mce())
250		return EPOLLIN | EPOLLRDNORM;
251	return 0;
252}
253
254static long mce_chrdev_ioctl(struct file *f, unsigned int cmd,
255				unsigned long arg)
256{
257	int __user *p = (int __user *)arg;
258
259	if (!capable(CAP_SYS_ADMIN))
260		return -EPERM;
261
262	switch (cmd) {
263	case MCE_GET_RECORD_LEN:
264		return put_user(sizeof(struct mce), p);
265	case MCE_GET_LOG_LEN:
266		return put_user(mcelog->len, p);
267	case MCE_GETCLEAR_FLAGS: {
268		unsigned flags;
269
270		do {
271			flags = mcelog->flags;
272		} while (cmpxchg(&mcelog->flags, flags, 0) != flags);
273
274		return put_user(flags, p);
275	}
276	default:
277		return -ENOTTY;
278	}
279}
280
281void mce_register_injector_chain(struct notifier_block *nb)
282{
283	blocking_notifier_chain_register(&mce_injector_chain, nb);
284}
285EXPORT_SYMBOL_GPL(mce_register_injector_chain);
286
287void mce_unregister_injector_chain(struct notifier_block *nb)
288{
289	blocking_notifier_chain_unregister(&mce_injector_chain, nb);
290}
291EXPORT_SYMBOL_GPL(mce_unregister_injector_chain);
292
293static ssize_t mce_chrdev_write(struct file *filp, const char __user *ubuf,
294				size_t usize, loff_t *off)
295{
296	struct mce m;
297
298	if (!capable(CAP_SYS_ADMIN))
299		return -EPERM;
300	/*
301	 * There are some cases where real MSR reads could slip
302	 * through.
303	 */
304	if (!boot_cpu_has(X86_FEATURE_MCE) || !boot_cpu_has(X86_FEATURE_MCA))
305		return -EIO;
306
307	if ((unsigned long)usize > sizeof(struct mce))
308		usize = sizeof(struct mce);
309	if (copy_from_user(&m, ubuf, usize))
310		return -EFAULT;
311
312	if (m.extcpu >= num_possible_cpus() || !cpu_online(m.extcpu))
313		return -EINVAL;
314
315	/*
316	 * Need to give user space some time to set everything up,
317	 * so do it a jiffie or two later everywhere.
318	 */
319	schedule_timeout(2);
320
321	blocking_notifier_call_chain(&mce_injector_chain, 0, &m);
322
323	return usize;
324}
325
326static const struct file_operations mce_chrdev_ops = {
327	.open			= mce_chrdev_open,
328	.release		= mce_chrdev_release,
329	.read			= mce_chrdev_read,
330	.write			= mce_chrdev_write,
331	.poll			= mce_chrdev_poll,
332	.unlocked_ioctl		= mce_chrdev_ioctl,
333	.compat_ioctl		= compat_ptr_ioctl,
334	.llseek			= no_llseek,
335};
336
337static struct miscdevice mce_chrdev_device = {
338	MISC_MCELOG_MINOR,
339	"mcelog",
340	&mce_chrdev_ops,
341};
342
343static __init int dev_mcelog_init_device(void)
344{
345	int mce_log_len;
346	int err;
347
348	mce_log_len = max(MCE_LOG_MIN_LEN, num_online_cpus());
349	mcelog = kzalloc(struct_size(mcelog, entry, mce_log_len), GFP_KERNEL);
350	if (!mcelog)
351		return -ENOMEM;
352
353	memcpy(mcelog->signature, MCE_LOG_SIGNATURE, sizeof(mcelog->signature));
354	mcelog->len = mce_log_len;
355	mcelog->recordlen = sizeof(struct mce);
356
357	/* register character device /dev/mcelog */
358	err = misc_register(&mce_chrdev_device);
359	if (err) {
360		if (err == -EBUSY)
361			/* Xen dom0 might have registered the device already. */
362			pr_info("Unable to init device /dev/mcelog, already registered");
363		else
364			pr_err("Unable to init device /dev/mcelog (rc: %d)\n", err);
365
366		kfree(mcelog);
367		return err;
368	}
369
370	mce_register_decode_chain(&dev_mcelog_nb);
371	return 0;
372}
373device_initcall_sync(dev_mcelog_init_device);
374