xref: /linux/arch/x86/kernel/cpu/mce/severity.c (revision ec6347bb43395cb92126788a1a5b25302543f815)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * MCE grading rules.
4  * Copyright 2008, 2009 Intel Corporation.
5  *
6  * Author: Andi Kleen
7  */
8 #include <linux/kernel.h>
9 #include <linux/seq_file.h>
10 #include <linux/init.h>
11 #include <linux/debugfs.h>
12 #include <linux/uaccess.h>
13 
14 #include <asm/mce.h>
15 #include <asm/intel-family.h>
16 
17 #include "internal.h"
18 
19 /*
20  * Grade an mce by severity. In general the most severe ones are processed
21  * first. Since there are quite a lot of combinations test the bits in a
22  * table-driven way. The rules are simply processed in order, first
23  * match wins.
24  *
25  * Note this is only used for machine check exceptions, the corrected
26  * errors use much simpler rules. The exceptions still check for the corrected
27  * errors, but only to leave them alone for the CMCI handler (except for
28  * panic situations)
29  */
30 
31 enum context { IN_KERNEL = 1, IN_USER = 2, IN_KERNEL_RECOV = 3 };
32 enum ser { SER_REQUIRED = 1, NO_SER = 2 };
33 enum exception { EXCP_CONTEXT = 1, NO_EXCP = 2 };
34 
35 static struct severity {
36 	u64 mask;
37 	u64 result;
38 	unsigned char sev;
39 	unsigned char mcgmask;
40 	unsigned char mcgres;
41 	unsigned char ser;
42 	unsigned char context;
43 	unsigned char excp;
44 	unsigned char covered;
45 	unsigned char cpu_model;
46 	unsigned char cpu_minstepping;
47 	unsigned char bank_lo, bank_hi;
48 	char *msg;
49 } severities[] = {
50 #define MCESEV(s, m, c...) { .sev = MCE_ ## s ## _SEVERITY, .msg = m, ## c }
51 #define BANK_RANGE(l, h) .bank_lo = l, .bank_hi = h
52 #define MODEL_STEPPING(m, s) .cpu_model = m, .cpu_minstepping = s
53 #define  KERNEL		.context = IN_KERNEL
54 #define  USER		.context = IN_USER
55 #define  KERNEL_RECOV	.context = IN_KERNEL_RECOV
56 #define  SER		.ser = SER_REQUIRED
57 #define  NOSER		.ser = NO_SER
58 #define  EXCP		.excp = EXCP_CONTEXT
59 #define  NOEXCP		.excp = NO_EXCP
60 #define  BITCLR(x)	.mask = x, .result = 0
61 #define  BITSET(x)	.mask = x, .result = x
62 #define  MCGMASK(x, y)	.mcgmask = x, .mcgres = y
63 #define  MASK(x, y)	.mask = x, .result = y
64 #define MCI_UC_S (MCI_STATUS_UC|MCI_STATUS_S)
65 #define MCI_UC_AR (MCI_STATUS_UC|MCI_STATUS_AR)
66 #define MCI_UC_SAR (MCI_STATUS_UC|MCI_STATUS_S|MCI_STATUS_AR)
67 #define	MCI_ADDR (MCI_STATUS_ADDRV|MCI_STATUS_MISCV)
68 
69 	MCESEV(
70 		NO, "Invalid",
71 		BITCLR(MCI_STATUS_VAL)
72 		),
73 	MCESEV(
74 		NO, "Not enabled",
75 		EXCP, BITCLR(MCI_STATUS_EN)
76 		),
77 	MCESEV(
78 		PANIC, "Processor context corrupt",
79 		BITSET(MCI_STATUS_PCC)
80 		),
81 	/* When MCIP is not set something is very confused */
82 	MCESEV(
83 		PANIC, "MCIP not set in MCA handler",
84 		EXCP, MCGMASK(MCG_STATUS_MCIP, 0)
85 		),
86 	/* Neither return not error IP -- no chance to recover -> PANIC */
87 	MCESEV(
88 		PANIC, "Neither restart nor error IP",
89 		EXCP, MCGMASK(MCG_STATUS_RIPV|MCG_STATUS_EIPV, 0)
90 		),
91 	MCESEV(
92 		PANIC, "In kernel and no restart IP",
93 		EXCP, KERNEL, MCGMASK(MCG_STATUS_RIPV, 0)
94 		),
95 	MCESEV(
96 		PANIC, "In kernel and no restart IP",
97 		EXCP, KERNEL_RECOV, MCGMASK(MCG_STATUS_RIPV, 0)
98 		),
99 	MCESEV(
100 		KEEP, "Corrected error",
101 		NOSER, BITCLR(MCI_STATUS_UC)
102 		),
103 	/*
104 	 * known AO MCACODs reported via MCE or CMC:
105 	 *
106 	 * SRAO could be signaled either via a machine check exception or
107 	 * CMCI with the corresponding bit S 1 or 0. So we don't need to
108 	 * check bit S for SRAO.
109 	 */
110 	MCESEV(
111 		AO, "Action optional: memory scrubbing error",
112 		SER, MASK(MCI_UC_AR|MCACOD_SCRUBMSK, MCI_STATUS_UC|MCACOD_SCRUB)
113 		),
114 	MCESEV(
115 		AO, "Action optional: last level cache writeback error",
116 		SER, MASK(MCI_UC_AR|MCACOD, MCI_STATUS_UC|MCACOD_L3WB)
117 		),
118 	/*
119 	 * Quirk for Skylake/Cascade Lake. Patrol scrubber may be configured
120 	 * to report uncorrected errors using CMCI with a special signature.
121 	 * UC=0, MSCOD=0x0010, MCACOD=binary(000X 0000 1100 XXXX) reported
122 	 * in one of the memory controller banks.
123 	 * Set severity to "AO" for same action as normal patrol scrub error.
124 	 */
125 	MCESEV(
126 		AO, "Uncorrected Patrol Scrub Error",
127 		SER, MASK(MCI_STATUS_UC|MCI_ADDR|0xffffeff0, MCI_ADDR|0x001000c0),
128 		MODEL_STEPPING(INTEL_FAM6_SKYLAKE_X, 4), BANK_RANGE(13, 18)
129 	),
130 
131 	/* ignore OVER for UCNA */
132 	MCESEV(
133 		UCNA, "Uncorrected no action required",
134 		SER, MASK(MCI_UC_SAR, MCI_STATUS_UC)
135 		),
136 	MCESEV(
137 		PANIC, "Illegal combination (UCNA with AR=1)",
138 		SER,
139 		MASK(MCI_STATUS_OVER|MCI_UC_SAR, MCI_STATUS_UC|MCI_STATUS_AR)
140 		),
141 	MCESEV(
142 		KEEP, "Non signalled machine check",
143 		SER, BITCLR(MCI_STATUS_S)
144 		),
145 
146 	MCESEV(
147 		PANIC, "Action required with lost events",
148 		SER, BITSET(MCI_STATUS_OVER|MCI_UC_SAR)
149 		),
150 
151 	/* known AR MCACODs: */
152 #ifdef	CONFIG_MEMORY_FAILURE
153 	MCESEV(
154 		KEEP, "Action required but unaffected thread is continuable",
155 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR|MCI_ADDR, MCI_UC_SAR|MCI_ADDR),
156 		MCGMASK(MCG_STATUS_RIPV|MCG_STATUS_EIPV, MCG_STATUS_RIPV)
157 		),
158 	MCESEV(
159 		AR, "Action required: data load in error recoverable area of kernel",
160 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR|MCI_ADDR|MCACOD, MCI_UC_SAR|MCI_ADDR|MCACOD_DATA),
161 		KERNEL_RECOV
162 		),
163 	MCESEV(
164 		AR, "Action required: data load error in a user process",
165 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR|MCI_ADDR|MCACOD, MCI_UC_SAR|MCI_ADDR|MCACOD_DATA),
166 		USER
167 		),
168 	MCESEV(
169 		AR, "Action required: instruction fetch error in a user process",
170 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR|MCI_ADDR|MCACOD, MCI_UC_SAR|MCI_ADDR|MCACOD_INSTR),
171 		USER
172 		),
173 	MCESEV(
174 		PANIC, "Data load in unrecoverable area of kernel",
175 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR|MCI_ADDR|MCACOD, MCI_UC_SAR|MCI_ADDR|MCACOD_DATA),
176 		KERNEL
177 		),
178 	MCESEV(
179 		PANIC, "Instruction fetch error in kernel",
180 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR|MCI_ADDR|MCACOD, MCI_UC_SAR|MCI_ADDR|MCACOD_INSTR),
181 		KERNEL
182 		),
183 #endif
184 	MCESEV(
185 		PANIC, "Action required: unknown MCACOD",
186 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR, MCI_UC_SAR)
187 		),
188 
189 	MCESEV(
190 		SOME, "Action optional: unknown MCACOD",
191 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR, MCI_UC_S)
192 		),
193 	MCESEV(
194 		SOME, "Action optional with lost events",
195 		SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR, MCI_STATUS_OVER|MCI_UC_S)
196 		),
197 
198 	MCESEV(
199 		PANIC, "Overflowed uncorrected",
200 		BITSET(MCI_STATUS_OVER|MCI_STATUS_UC)
201 		),
202 	MCESEV(
203 		UC, "Uncorrected",
204 		BITSET(MCI_STATUS_UC)
205 		),
206 	MCESEV(
207 		SOME, "No match",
208 		BITSET(0)
209 		)	/* always matches. keep at end */
210 };
211 
212 #define mc_recoverable(mcg) (((mcg) & (MCG_STATUS_RIPV|MCG_STATUS_EIPV)) == \
213 				(MCG_STATUS_RIPV|MCG_STATUS_EIPV))
214 
215 /*
216  * If mcgstatus indicated that ip/cs on the stack were
217  * no good, then "m->cs" will be zero and we will have
218  * to assume the worst case (IN_KERNEL) as we actually
219  * have no idea what we were executing when the machine
220  * check hit.
221  * If we do have a good "m->cs" (or a faked one in the
222  * case we were executing in VM86 mode) we can use it to
223  * distinguish an exception taken in user from from one
224  * taken in the kernel.
225  */
226 static int error_context(struct mce *m)
227 {
228 	if ((m->cs & 3) == 3)
229 		return IN_USER;
230 
231 	if (mc_recoverable(m->mcgstatus) && ex_has_fault_handler(m->ip)) {
232 		m->kflags |= MCE_IN_KERNEL_RECOV;
233 		return IN_KERNEL_RECOV;
234 	}
235 
236 	return IN_KERNEL;
237 }
238 
239 static int mce_severity_amd_smca(struct mce *m, enum context err_ctx)
240 {
241 	u32 addr = MSR_AMD64_SMCA_MCx_CONFIG(m->bank);
242 	u32 low, high;
243 
244 	/*
245 	 * We need to look at the following bits:
246 	 * - "succor" bit (data poisoning support), and
247 	 * - TCC bit (Task Context Corrupt)
248 	 * in MCi_STATUS to determine error severity.
249 	 */
250 	if (!mce_flags.succor)
251 		return MCE_PANIC_SEVERITY;
252 
253 	if (rdmsr_safe(addr, &low, &high))
254 		return MCE_PANIC_SEVERITY;
255 
256 	/* TCC (Task context corrupt). If set and if IN_KERNEL, panic. */
257 	if ((low & MCI_CONFIG_MCAX) &&
258 	    (m->status & MCI_STATUS_TCC) &&
259 	    (err_ctx == IN_KERNEL))
260 		return MCE_PANIC_SEVERITY;
261 
262 	 /* ...otherwise invoke hwpoison handler. */
263 	return MCE_AR_SEVERITY;
264 }
265 
266 /*
267  * See AMD Error Scope Hierarchy table in a newer BKDG. For example
268  * 49125_15h_Models_30h-3Fh_BKDG.pdf, section "RAS Features"
269  */
270 static int mce_severity_amd(struct mce *m, int tolerant, char **msg, bool is_excp)
271 {
272 	enum context ctx = error_context(m);
273 
274 	/* Processor Context Corrupt, no need to fumble too much, die! */
275 	if (m->status & MCI_STATUS_PCC)
276 		return MCE_PANIC_SEVERITY;
277 
278 	if (m->status & MCI_STATUS_UC) {
279 
280 		if (ctx == IN_KERNEL)
281 			return MCE_PANIC_SEVERITY;
282 
283 		/*
284 		 * On older systems where overflow_recov flag is not present, we
285 		 * should simply panic if an error overflow occurs. If
286 		 * overflow_recov flag is present and set, then software can try
287 		 * to at least kill process to prolong system operation.
288 		 */
289 		if (mce_flags.overflow_recov) {
290 			if (mce_flags.smca)
291 				return mce_severity_amd_smca(m, ctx);
292 
293 			/* kill current process */
294 			return MCE_AR_SEVERITY;
295 		} else {
296 			/* at least one error was not logged */
297 			if (m->status & MCI_STATUS_OVER)
298 				return MCE_PANIC_SEVERITY;
299 		}
300 
301 		/*
302 		 * For any other case, return MCE_UC_SEVERITY so that we log the
303 		 * error and exit #MC handler.
304 		 */
305 		return MCE_UC_SEVERITY;
306 	}
307 
308 	/*
309 	 * deferred error: poll handler catches these and adds to mce_ring so
310 	 * memory-failure can take recovery actions.
311 	 */
312 	if (m->status & MCI_STATUS_DEFERRED)
313 		return MCE_DEFERRED_SEVERITY;
314 
315 	/*
316 	 * corrected error: poll handler catches these and passes responsibility
317 	 * of decoding the error to EDAC
318 	 */
319 	return MCE_KEEP_SEVERITY;
320 }
321 
322 static int mce_severity_intel(struct mce *m, int tolerant, char **msg, bool is_excp)
323 {
324 	enum exception excp = (is_excp ? EXCP_CONTEXT : NO_EXCP);
325 	enum context ctx = error_context(m);
326 	struct severity *s;
327 
328 	for (s = severities;; s++) {
329 		if ((m->status & s->mask) != s->result)
330 			continue;
331 		if ((m->mcgstatus & s->mcgmask) != s->mcgres)
332 			continue;
333 		if (s->ser == SER_REQUIRED && !mca_cfg.ser)
334 			continue;
335 		if (s->ser == NO_SER && mca_cfg.ser)
336 			continue;
337 		if (s->context && ctx != s->context)
338 			continue;
339 		if (s->excp && excp != s->excp)
340 			continue;
341 		if (s->cpu_model && boot_cpu_data.x86_model != s->cpu_model)
342 			continue;
343 		if (s->cpu_minstepping && boot_cpu_data.x86_stepping < s->cpu_minstepping)
344 			continue;
345 		if (s->bank_lo && (m->bank < s->bank_lo || m->bank > s->bank_hi))
346 			continue;
347 		if (msg)
348 			*msg = s->msg;
349 		s->covered = 1;
350 		if (s->sev >= MCE_UC_SEVERITY && ctx == IN_KERNEL) {
351 			if (tolerant < 1)
352 				return MCE_PANIC_SEVERITY;
353 		}
354 		return s->sev;
355 	}
356 }
357 
358 /* Default to mce_severity_intel */
359 int (*mce_severity)(struct mce *m, int tolerant, char **msg, bool is_excp) =
360 		    mce_severity_intel;
361 
362 void __init mcheck_vendor_init_severity(void)
363 {
364 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
365 	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
366 		mce_severity = mce_severity_amd;
367 }
368 
369 #ifdef CONFIG_DEBUG_FS
370 static void *s_start(struct seq_file *f, loff_t *pos)
371 {
372 	if (*pos >= ARRAY_SIZE(severities))
373 		return NULL;
374 	return &severities[*pos];
375 }
376 
377 static void *s_next(struct seq_file *f, void *data, loff_t *pos)
378 {
379 	if (++(*pos) >= ARRAY_SIZE(severities))
380 		return NULL;
381 	return &severities[*pos];
382 }
383 
384 static void s_stop(struct seq_file *f, void *data)
385 {
386 }
387 
388 static int s_show(struct seq_file *f, void *data)
389 {
390 	struct severity *ser = data;
391 	seq_printf(f, "%d\t%s\n", ser->covered, ser->msg);
392 	return 0;
393 }
394 
395 static const struct seq_operations severities_seq_ops = {
396 	.start	= s_start,
397 	.next	= s_next,
398 	.stop	= s_stop,
399 	.show	= s_show,
400 };
401 
402 static int severities_coverage_open(struct inode *inode, struct file *file)
403 {
404 	return seq_open(file, &severities_seq_ops);
405 }
406 
407 static ssize_t severities_coverage_write(struct file *file,
408 					 const char __user *ubuf,
409 					 size_t count, loff_t *ppos)
410 {
411 	int i;
412 	for (i = 0; i < ARRAY_SIZE(severities); i++)
413 		severities[i].covered = 0;
414 	return count;
415 }
416 
417 static const struct file_operations severities_coverage_fops = {
418 	.open		= severities_coverage_open,
419 	.release	= seq_release,
420 	.read		= seq_read,
421 	.write		= severities_coverage_write,
422 	.llseek		= seq_lseek,
423 };
424 
425 static int __init severities_debugfs_init(void)
426 {
427 	struct dentry *dmce;
428 
429 	dmce = mce_get_debugfs_dir();
430 
431 	debugfs_create_file("severities-coverage", 0444, dmce, NULL,
432 			    &severities_coverage_fops);
433 	return 0;
434 }
435 late_initcall(severities_debugfs_init);
436 #endif /* CONFIG_DEBUG_FS */
437