1/*
2 * Machine check injection support.
3 * Copyright 2008 Intel Corporation.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License
7 * as published by the Free Software Foundation; version 2
8 * of the License.
9 *
10 * Authors:
11 * Andi Kleen
12 * Ying Huang
13 */
14#include <linux/uaccess.h>
15#include <linux/module.h>
16#include <linux/timer.h>
17#include <linux/kernel.h>
18#include <linux/string.h>
19#include <linux/fs.h>
20#include <linux/preempt.h>
21#include <linux/smp.h>
22#include <linux/notifier.h>
23#include <linux/kdebug.h>
24#include <linux/cpu.h>
25#include <linux/sched.h>
26#include <linux/gfp.h>
27#include <asm/mce.h>
28#include <asm/apic.h>
29#include <asm/nmi.h>
30
31/* Update fake mce registers on current CPU. */
32static void inject_mce(struct mce *m)
33{
34	struct mce *i = &per_cpu(injectm, m->extcpu);
35
36	/* Make sure no one reads partially written injectm */
37	i->finished = 0;
38	mb();
39	m->finished = 0;
40	/* First set the fields after finished */
41	i->extcpu = m->extcpu;
42	mb();
43	/* Now write record in order, finished last (except above) */
44	memcpy(i, m, sizeof(struct mce));
45	/* Finally activate it */
46	mb();
47	i->finished = 1;
48}
49
50static void raise_poll(struct mce *m)
51{
52	unsigned long flags;
53	mce_banks_t b;
54
55	memset(&b, 0xff, sizeof(mce_banks_t));
56	local_irq_save(flags);
57	machine_check_poll(0, &b);
58	local_irq_restore(flags);
59	m->finished = 0;
60}
61
62static void raise_exception(struct mce *m, struct pt_regs *pregs)
63{
64	struct pt_regs regs;
65	unsigned long flags;
66
67	if (!pregs) {
68		memset(&regs, 0, sizeof(struct pt_regs));
69		regs.ip = m->ip;
70		regs.cs = m->cs;
71		pregs = &regs;
72	}
73	/* in mcheck exeception handler, irq will be disabled */
74	local_irq_save(flags);
75	do_machine_check(pregs, 0);
76	local_irq_restore(flags);
77	m->finished = 0;
78}
79
80static cpumask_var_t mce_inject_cpumask;
81static DEFINE_MUTEX(mce_inject_mutex);
82
83static int mce_raise_notify(unsigned int cmd, struct pt_regs *regs)
84{
85	int cpu = smp_processor_id();
86	struct mce *m = this_cpu_ptr(&injectm);
87	if (!cpumask_test_cpu(cpu, mce_inject_cpumask))
88		return NMI_DONE;
89	cpumask_clear_cpu(cpu, mce_inject_cpumask);
90	if (m->inject_flags & MCJ_EXCEPTION)
91		raise_exception(m, regs);
92	else if (m->status)
93		raise_poll(m);
94	return NMI_HANDLED;
95}
96
97static void mce_irq_ipi(void *info)
98{
99	int cpu = smp_processor_id();
100	struct mce *m = this_cpu_ptr(&injectm);
101
102	if (cpumask_test_cpu(cpu, mce_inject_cpumask) &&
103			m->inject_flags & MCJ_EXCEPTION) {
104		cpumask_clear_cpu(cpu, mce_inject_cpumask);
105		raise_exception(m, NULL);
106	}
107}
108
109/* Inject mce on current CPU */
110static int raise_local(void)
111{
112	struct mce *m = this_cpu_ptr(&injectm);
113	int context = MCJ_CTX(m->inject_flags);
114	int ret = 0;
115	int cpu = m->extcpu;
116
117	if (m->inject_flags & MCJ_EXCEPTION) {
118		printk(KERN_INFO "Triggering MCE exception on CPU %d\n", cpu);
119		switch (context) {
120		case MCJ_CTX_IRQ:
121			/*
122			 * Could do more to fake interrupts like
123			 * calling irq_enter, but the necessary
124			 * machinery isn't exported currently.
125			 */
126			/*FALL THROUGH*/
127		case MCJ_CTX_PROCESS:
128			raise_exception(m, NULL);
129			break;
130		default:
131			printk(KERN_INFO "Invalid MCE context\n");
132			ret = -EINVAL;
133		}
134		printk(KERN_INFO "MCE exception done on CPU %d\n", cpu);
135	} else if (m->status) {
136		printk(KERN_INFO "Starting machine check poll CPU %d\n", cpu);
137		raise_poll(m);
138		mce_notify_irq();
139		printk(KERN_INFO "Machine check poll done on CPU %d\n", cpu);
140	} else
141		m->finished = 0;
142
143	return ret;
144}
145
146static void raise_mce(struct mce *m)
147{
148	int context = MCJ_CTX(m->inject_flags);
149
150	inject_mce(m);
151
152	if (context == MCJ_CTX_RANDOM)
153		return;
154
155#ifdef CONFIG_X86_LOCAL_APIC
156	if (m->inject_flags & (MCJ_IRQ_BROADCAST | MCJ_NMI_BROADCAST)) {
157		unsigned long start;
158		int cpu;
159
160		get_online_cpus();
161		cpumask_copy(mce_inject_cpumask, cpu_online_mask);
162		cpumask_clear_cpu(get_cpu(), mce_inject_cpumask);
163		for_each_online_cpu(cpu) {
164			struct mce *mcpu = &per_cpu(injectm, cpu);
165			if (!mcpu->finished ||
166			    MCJ_CTX(mcpu->inject_flags) != MCJ_CTX_RANDOM)
167				cpumask_clear_cpu(cpu, mce_inject_cpumask);
168		}
169		if (!cpumask_empty(mce_inject_cpumask)) {
170			if (m->inject_flags & MCJ_IRQ_BROADCAST) {
171				/*
172				 * don't wait because mce_irq_ipi is necessary
173				 * to be sync with following raise_local
174				 */
175				preempt_disable();
176				smp_call_function_many(mce_inject_cpumask,
177					mce_irq_ipi, NULL, 0);
178				preempt_enable();
179			} else if (m->inject_flags & MCJ_NMI_BROADCAST)
180				apic->send_IPI_mask(mce_inject_cpumask,
181						NMI_VECTOR);
182		}
183		start = jiffies;
184		while (!cpumask_empty(mce_inject_cpumask)) {
185			if (!time_before(jiffies, start + 2*HZ)) {
186				printk(KERN_ERR
187				"Timeout waiting for mce inject %lx\n",
188					*cpumask_bits(mce_inject_cpumask));
189				break;
190			}
191			cpu_relax();
192		}
193		raise_local();
194		put_cpu();
195		put_online_cpus();
196	} else
197#endif
198	{
199		preempt_disable();
200		raise_local();
201		preempt_enable();
202	}
203}
204
205/* Error injection interface */
206static ssize_t mce_write(struct file *filp, const char __user *ubuf,
207			 size_t usize, loff_t *off)
208{
209	struct mce m;
210
211	if (!capable(CAP_SYS_ADMIN))
212		return -EPERM;
213	/*
214	 * There are some cases where real MSR reads could slip
215	 * through.
216	 */
217	if (!boot_cpu_has(X86_FEATURE_MCE) || !boot_cpu_has(X86_FEATURE_MCA))
218		return -EIO;
219
220	if ((unsigned long)usize > sizeof(struct mce))
221		usize = sizeof(struct mce);
222	if (copy_from_user(&m, ubuf, usize))
223		return -EFAULT;
224
225	if (m.extcpu >= num_possible_cpus() || !cpu_online(m.extcpu))
226		return -EINVAL;
227
228	/*
229	 * Need to give user space some time to set everything up,
230	 * so do it a jiffie or two later everywhere.
231	 */
232	schedule_timeout(2);
233
234	mutex_lock(&mce_inject_mutex);
235	raise_mce(&m);
236	mutex_unlock(&mce_inject_mutex);
237	return usize;
238}
239
240static int inject_init(void)
241{
242	if (!alloc_cpumask_var(&mce_inject_cpumask, GFP_KERNEL))
243		return -ENOMEM;
244	printk(KERN_INFO "Machine check injector initialized\n");
245	register_mce_write_callback(mce_write);
246	register_nmi_handler(NMI_LOCAL, mce_raise_notify, 0,
247				"mce_notify");
248	return 0;
249}
250
251module_init(inject_init);
252/*
253 * Cannot tolerate unloading currently because we cannot
254 * guarantee all openers of mce_chrdev will get a reference to us.
255 */
256MODULE_LICENSE("GPL");
257