1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Intel specific MCE features.
4  * Copyright 2004 Zwane Mwaikambo <zwane@linuxpower.ca>
5  * Copyright (C) 2008, 2009 Intel Corporation
6  * Author: Andi Kleen
7  */
8 
9 #include <linux/gfp.h>
10 #include <linux/interrupt.h>
11 #include <linux/percpu.h>
12 #include <linux/sched.h>
13 #include <linux/cpumask.h>
14 #include <asm/apic.h>
15 #include <asm/cpufeature.h>
16 #include <asm/intel-family.h>
17 #include <asm/processor.h>
18 #include <asm/msr.h>
19 #include <asm/mce.h>
20 
21 #include "internal.h"
22 
23 /*
24  * Support for Intel Correct Machine Check Interrupts. This allows
25  * the CPU to raise an interrupt when a corrected machine check happened.
26  * Normally we pick those up using a regular polling timer.
27  * Also supports reliable discovery of shared banks.
28  */
29 
30 /*
31  * CMCI can be delivered to multiple cpus that share a machine check bank
32  * so we need to designate a single cpu to process errors logged in each bank
33  * in the interrupt handler (otherwise we would have many races and potential
34  * double reporting of the same error).
35  * Note that this can change when a cpu is offlined or brought online since
36  * some MCA banks are shared across cpus. When a cpu is offlined, cmci_clear()
37  * disables CMCI on all banks owned by the cpu and clears this bitfield. At
38  * this point, cmci_rediscover() kicks in and a different cpu may end up
39  * taking ownership of some of the shared MCA banks that were previously
40  * owned by the offlined cpu.
41  */
42 static DEFINE_PER_CPU(mce_banks_t, mce_banks_owned);
43 
44 /*
45  * CMCI storm detection backoff counter
46  *
47  * During storm, we reset this counter to INITIAL_CHECK_INTERVAL in case we've
48  * encountered an error. If not, we decrement it by one. We signal the end of
49  * the CMCI storm when it reaches 0.
50  */
51 static DEFINE_PER_CPU(int, cmci_backoff_cnt);
52 
53 /*
54  * cmci_discover_lock protects against parallel discovery attempts
55  * which could race against each other.
56  */
57 static DEFINE_RAW_SPINLOCK(cmci_discover_lock);
58 
59 #define CMCI_THRESHOLD		1
60 #define CMCI_POLL_INTERVAL	(30 * HZ)
61 #define CMCI_STORM_INTERVAL	(HZ)
62 #define CMCI_STORM_THRESHOLD	15
63 
64 static DEFINE_PER_CPU(unsigned long, cmci_time_stamp);
65 static DEFINE_PER_CPU(unsigned int, cmci_storm_cnt);
66 static DEFINE_PER_CPU(unsigned int, cmci_storm_state);
67 
68 enum {
69 	CMCI_STORM_NONE,
70 	CMCI_STORM_ACTIVE,
71 	CMCI_STORM_SUBSIDED,
72 };
73 
74 static atomic_t cmci_storm_on_cpus;
75 
cmci_supported(int * banks)76 static int cmci_supported(int *banks)
77 {
78 	u64 cap;
79 
80 	if (mca_cfg.cmci_disabled || mca_cfg.ignore_ce)
81 		return 0;
82 
83 	/*
84 	 * Vendor check is not strictly needed, but the initial
85 	 * initialization is vendor keyed and this
86 	 * makes sure none of the backdoors are entered otherwise.
87 	 */
88 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL &&
89 	    boot_cpu_data.x86_vendor != X86_VENDOR_ZHAOXIN)
90 		return 0;
91 
92 	if (!boot_cpu_has(X86_FEATURE_APIC) || lapic_get_maxlvt() < 6)
93 		return 0;
94 	rdmsrl(MSR_IA32_MCG_CAP, cap);
95 	*banks = min_t(unsigned, MAX_NR_BANKS, cap & 0xff);
96 	return !!(cap & MCG_CMCI_P);
97 }
98 
lmce_supported(void)99 static bool lmce_supported(void)
100 {
101 	u64 tmp;
102 
103 	if (mca_cfg.lmce_disabled)
104 		return false;
105 
106 	rdmsrl(MSR_IA32_MCG_CAP, tmp);
107 
108 	/*
109 	 * LMCE depends on recovery support in the processor. Hence both
110 	 * MCG_SER_P and MCG_LMCE_P should be present in MCG_CAP.
111 	 */
112 	if ((tmp & (MCG_SER_P | MCG_LMCE_P)) !=
113 		   (MCG_SER_P | MCG_LMCE_P))
114 		return false;
115 
116 	/*
117 	 * BIOS should indicate support for LMCE by setting bit 20 in
118 	 * IA32_FEAT_CTL without which touching MCG_EXT_CTL will generate a #GP
119 	 * fault.  The MSR must also be locked for LMCE_ENABLED to take effect.
120 	 * WARN if the MSR isn't locked as init_ia32_feat_ctl() unconditionally
121 	 * locks the MSR in the event that it wasn't already locked by BIOS.
122 	 */
123 	rdmsrl(MSR_IA32_FEAT_CTL, tmp);
124 	if (WARN_ON_ONCE(!(tmp & FEAT_CTL_LOCKED)))
125 		return false;
126 
127 	return tmp & FEAT_CTL_LMCE_ENABLED;
128 }
129 
mce_intel_cmci_poll(void)130 bool mce_intel_cmci_poll(void)
131 {
132 	if (__this_cpu_read(cmci_storm_state) == CMCI_STORM_NONE)
133 		return false;
134 
135 	/*
136 	 * Reset the counter if we've logged an error in the last poll
137 	 * during the storm.
138 	 */
139 	if (machine_check_poll(0, this_cpu_ptr(&mce_banks_owned)))
140 		this_cpu_write(cmci_backoff_cnt, INITIAL_CHECK_INTERVAL);
141 	else
142 		this_cpu_dec(cmci_backoff_cnt);
143 
144 	return true;
145 }
146 
mce_intel_hcpu_update(unsigned long cpu)147 void mce_intel_hcpu_update(unsigned long cpu)
148 {
149 	if (per_cpu(cmci_storm_state, cpu) == CMCI_STORM_ACTIVE)
150 		atomic_dec(&cmci_storm_on_cpus);
151 
152 	per_cpu(cmci_storm_state, cpu) = CMCI_STORM_NONE;
153 }
154 
cmci_toggle_interrupt_mode(bool on)155 static void cmci_toggle_interrupt_mode(bool on)
156 {
157 	unsigned long flags, *owned;
158 	int bank;
159 	u64 val;
160 
161 	raw_spin_lock_irqsave(&cmci_discover_lock, flags);
162 	owned = this_cpu_ptr(mce_banks_owned);
163 	for_each_set_bit(bank, owned, MAX_NR_BANKS) {
164 		rdmsrl(MSR_IA32_MCx_CTL2(bank), val);
165 
166 		if (on)
167 			val |= MCI_CTL2_CMCI_EN;
168 		else
169 			val &= ~MCI_CTL2_CMCI_EN;
170 
171 		wrmsrl(MSR_IA32_MCx_CTL2(bank), val);
172 	}
173 	raw_spin_unlock_irqrestore(&cmci_discover_lock, flags);
174 }
175 
cmci_intel_adjust_timer(unsigned long interval)176 unsigned long cmci_intel_adjust_timer(unsigned long interval)
177 {
178 	if ((this_cpu_read(cmci_backoff_cnt) > 0) &&
179 	    (__this_cpu_read(cmci_storm_state) == CMCI_STORM_ACTIVE)) {
180 		mce_notify_irq();
181 		return CMCI_STORM_INTERVAL;
182 	}
183 
184 	switch (__this_cpu_read(cmci_storm_state)) {
185 	case CMCI_STORM_ACTIVE:
186 
187 		/*
188 		 * We switch back to interrupt mode once the poll timer has
189 		 * silenced itself. That means no events recorded and the timer
190 		 * interval is back to our poll interval.
191 		 */
192 		__this_cpu_write(cmci_storm_state, CMCI_STORM_SUBSIDED);
193 		if (!atomic_sub_return(1, &cmci_storm_on_cpus))
194 			pr_notice("CMCI storm subsided: switching to interrupt mode\n");
195 
196 		fallthrough;
197 
198 	case CMCI_STORM_SUBSIDED:
199 		/*
200 		 * We wait for all CPUs to go back to SUBSIDED state. When that
201 		 * happens we switch back to interrupt mode.
202 		 */
203 		if (!atomic_read(&cmci_storm_on_cpus)) {
204 			__this_cpu_write(cmci_storm_state, CMCI_STORM_NONE);
205 			cmci_toggle_interrupt_mode(true);
206 			cmci_recheck();
207 		}
208 		return CMCI_POLL_INTERVAL;
209 	default:
210 
211 		/* We have shiny weather. Let the poll do whatever it thinks. */
212 		return interval;
213 	}
214 }
215 
cmci_storm_detect(void)216 static bool cmci_storm_detect(void)
217 {
218 	unsigned int cnt = __this_cpu_read(cmci_storm_cnt);
219 	unsigned long ts = __this_cpu_read(cmci_time_stamp);
220 	unsigned long now = jiffies;
221 	int r;
222 
223 	if (__this_cpu_read(cmci_storm_state) != CMCI_STORM_NONE)
224 		return true;
225 
226 	if (time_before_eq(now, ts + CMCI_STORM_INTERVAL)) {
227 		cnt++;
228 	} else {
229 		cnt = 1;
230 		__this_cpu_write(cmci_time_stamp, now);
231 	}
232 	__this_cpu_write(cmci_storm_cnt, cnt);
233 
234 	if (cnt <= CMCI_STORM_THRESHOLD)
235 		return false;
236 
237 	cmci_toggle_interrupt_mode(false);
238 	__this_cpu_write(cmci_storm_state, CMCI_STORM_ACTIVE);
239 	r = atomic_add_return(1, &cmci_storm_on_cpus);
240 	mce_timer_kick(CMCI_STORM_INTERVAL);
241 	this_cpu_write(cmci_backoff_cnt, INITIAL_CHECK_INTERVAL);
242 
243 	if (r == 1)
244 		pr_notice("CMCI storm detected: switching to poll mode\n");
245 	return true;
246 }
247 
248 /*
249  * The interrupt handler. This is called on every event.
250  * Just call the poller directly to log any events.
251  * This could in theory increase the threshold under high load,
252  * but doesn't for now.
253  */
intel_threshold_interrupt(void)254 static void intel_threshold_interrupt(void)
255 {
256 	if (cmci_storm_detect())
257 		return;
258 
259 	machine_check_poll(MCP_TIMESTAMP, this_cpu_ptr(&mce_banks_owned));
260 }
261 
262 /*
263  * Enable CMCI (Corrected Machine Check Interrupt) for available MCE banks
264  * on this CPU. Use the algorithm recommended in the SDM to discover shared
265  * banks.
266  */
cmci_discover(int banks)267 static void cmci_discover(int banks)
268 {
269 	unsigned long *owned = (void *)this_cpu_ptr(&mce_banks_owned);
270 	unsigned long flags;
271 	int i;
272 	int bios_wrong_thresh = 0;
273 
274 	raw_spin_lock_irqsave(&cmci_discover_lock, flags);
275 	for (i = 0; i < banks; i++) {
276 		u64 val;
277 		int bios_zero_thresh = 0;
278 
279 		if (test_bit(i, owned))
280 			continue;
281 
282 		/* Skip banks in firmware first mode */
283 		if (test_bit(i, mce_banks_ce_disabled))
284 			continue;
285 
286 		rdmsrl(MSR_IA32_MCx_CTL2(i), val);
287 
288 		/* Already owned by someone else? */
289 		if (val & MCI_CTL2_CMCI_EN) {
290 			clear_bit(i, owned);
291 			__clear_bit(i, this_cpu_ptr(mce_poll_banks));
292 			continue;
293 		}
294 
295 		if (!mca_cfg.bios_cmci_threshold) {
296 			val &= ~MCI_CTL2_CMCI_THRESHOLD_MASK;
297 			val |= CMCI_THRESHOLD;
298 		} else if (!(val & MCI_CTL2_CMCI_THRESHOLD_MASK)) {
299 			/*
300 			 * If bios_cmci_threshold boot option was specified
301 			 * but the threshold is zero, we'll try to initialize
302 			 * it to 1.
303 			 */
304 			bios_zero_thresh = 1;
305 			val |= CMCI_THRESHOLD;
306 		}
307 
308 		val |= MCI_CTL2_CMCI_EN;
309 		wrmsrl(MSR_IA32_MCx_CTL2(i), val);
310 		rdmsrl(MSR_IA32_MCx_CTL2(i), val);
311 
312 		/* Did the enable bit stick? -- the bank supports CMCI */
313 		if (val & MCI_CTL2_CMCI_EN) {
314 			set_bit(i, owned);
315 			__clear_bit(i, this_cpu_ptr(mce_poll_banks));
316 			/*
317 			 * We are able to set thresholds for some banks that
318 			 * had a threshold of 0. This means the BIOS has not
319 			 * set the thresholds properly or does not work with
320 			 * this boot option. Note down now and report later.
321 			 */
322 			if (mca_cfg.bios_cmci_threshold && bios_zero_thresh &&
323 					(val & MCI_CTL2_CMCI_THRESHOLD_MASK))
324 				bios_wrong_thresh = 1;
325 		} else {
326 			WARN_ON(!test_bit(i, this_cpu_ptr(mce_poll_banks)));
327 		}
328 	}
329 	raw_spin_unlock_irqrestore(&cmci_discover_lock, flags);
330 	if (mca_cfg.bios_cmci_threshold && bios_wrong_thresh) {
331 		pr_info_once(
332 			"bios_cmci_threshold: Some banks do not have valid thresholds set\n");
333 		pr_info_once(
334 			"bios_cmci_threshold: Make sure your BIOS supports this boot option\n");
335 	}
336 }
337 
338 /*
339  * Just in case we missed an event during initialization check
340  * all the CMCI owned banks.
341  */
cmci_recheck(void)342 void cmci_recheck(void)
343 {
344 	unsigned long flags;
345 	int banks;
346 
347 	if (!mce_available(raw_cpu_ptr(&cpu_info)) || !cmci_supported(&banks))
348 		return;
349 
350 	local_irq_save(flags);
351 	machine_check_poll(0, this_cpu_ptr(&mce_banks_owned));
352 	local_irq_restore(flags);
353 }
354 
355 /* Caller must hold the lock on cmci_discover_lock */
__cmci_disable_bank(int bank)356 static void __cmci_disable_bank(int bank)
357 {
358 	u64 val;
359 
360 	if (!test_bit(bank, this_cpu_ptr(mce_banks_owned)))
361 		return;
362 	rdmsrl(MSR_IA32_MCx_CTL2(bank), val);
363 	val &= ~MCI_CTL2_CMCI_EN;
364 	wrmsrl(MSR_IA32_MCx_CTL2(bank), val);
365 	__clear_bit(bank, this_cpu_ptr(mce_banks_owned));
366 }
367 
368 /*
369  * Disable CMCI on this CPU for all banks it owns when it goes down.
370  * This allows other CPUs to claim the banks on rediscovery.
371  */
cmci_clear(void)372 void cmci_clear(void)
373 {
374 	unsigned long flags;
375 	int i;
376 	int banks;
377 
378 	if (!cmci_supported(&banks))
379 		return;
380 	raw_spin_lock_irqsave(&cmci_discover_lock, flags);
381 	for (i = 0; i < banks; i++)
382 		__cmci_disable_bank(i);
383 	raw_spin_unlock_irqrestore(&cmci_discover_lock, flags);
384 }
385 
cmci_rediscover_work_func(void * arg)386 static void cmci_rediscover_work_func(void *arg)
387 {
388 	int banks;
389 
390 	/* Recheck banks in case CPUs don't all have the same */
391 	if (cmci_supported(&banks))
392 		cmci_discover(banks);
393 }
394 
395 /* After a CPU went down cycle through all the others and rediscover */
cmci_rediscover(void)396 void cmci_rediscover(void)
397 {
398 	int banks;
399 
400 	if (!cmci_supported(&banks))
401 		return;
402 
403 	on_each_cpu(cmci_rediscover_work_func, NULL, 1);
404 }
405 
406 /*
407  * Reenable CMCI on this CPU in case a CPU down failed.
408  */
cmci_reenable(void)409 void cmci_reenable(void)
410 {
411 	int banks;
412 	if (cmci_supported(&banks))
413 		cmci_discover(banks);
414 }
415 
cmci_disable_bank(int bank)416 void cmci_disable_bank(int bank)
417 {
418 	int banks;
419 	unsigned long flags;
420 
421 	if (!cmci_supported(&banks))
422 		return;
423 
424 	raw_spin_lock_irqsave(&cmci_discover_lock, flags);
425 	__cmci_disable_bank(bank);
426 	raw_spin_unlock_irqrestore(&cmci_discover_lock, flags);
427 }
428 
intel_init_cmci(void)429 void intel_init_cmci(void)
430 {
431 	int banks;
432 
433 	if (!cmci_supported(&banks))
434 		return;
435 
436 	mce_threshold_vector = intel_threshold_interrupt;
437 	cmci_discover(banks);
438 	/*
439 	 * For CPU #0 this runs with still disabled APIC, but that's
440 	 * ok because only the vector is set up. We still do another
441 	 * check for the banks later for CPU #0 just to make sure
442 	 * to not miss any events.
443 	 */
444 	apic_write(APIC_LVTCMCI, THRESHOLD_APIC_VECTOR|APIC_DM_FIXED);
445 	cmci_recheck();
446 }
447 
intel_init_lmce(void)448 void intel_init_lmce(void)
449 {
450 	u64 val;
451 
452 	if (!lmce_supported())
453 		return;
454 
455 	rdmsrl(MSR_IA32_MCG_EXT_CTL, val);
456 
457 	if (!(val & MCG_EXT_CTL_LMCE_EN))
458 		wrmsrl(MSR_IA32_MCG_EXT_CTL, val | MCG_EXT_CTL_LMCE_EN);
459 }
460 
intel_clear_lmce(void)461 void intel_clear_lmce(void)
462 {
463 	u64 val;
464 
465 	if (!lmce_supported())
466 		return;
467 
468 	rdmsrl(MSR_IA32_MCG_EXT_CTL, val);
469 	val &= ~MCG_EXT_CTL_LMCE_EN;
470 	wrmsrl(MSR_IA32_MCG_EXT_CTL, val);
471 }
472 
473 /*
474  * Enable additional error logs from the integrated
475  * memory controller on processors that support this.
476  */
intel_imc_init(struct cpuinfo_x86 * c)477 static void intel_imc_init(struct cpuinfo_x86 *c)
478 {
479 	u64 error_control;
480 
481 	switch (c->x86_model) {
482 	case INTEL_FAM6_SANDYBRIDGE_X:
483 	case INTEL_FAM6_IVYBRIDGE_X:
484 	case INTEL_FAM6_HASWELL_X:
485 		if (rdmsrl_safe(MSR_ERROR_CONTROL, &error_control))
486 			return;
487 		error_control |= 2;
488 		wrmsrl_safe(MSR_ERROR_CONTROL, error_control);
489 		break;
490 	}
491 }
492 
mce_intel_feature_init(struct cpuinfo_x86 * c)493 void mce_intel_feature_init(struct cpuinfo_x86 *c)
494 {
495 	intel_init_cmci();
496 	intel_init_lmce();
497 	intel_imc_init(c);
498 }
499 
mce_intel_feature_clear(struct cpuinfo_x86 * c)500 void mce_intel_feature_clear(struct cpuinfo_x86 *c)
501 {
502 	intel_clear_lmce();
503 }
504 
intel_filter_mce(struct mce * m)505 bool intel_filter_mce(struct mce *m)
506 {
507 	struct cpuinfo_x86 *c = &boot_cpu_data;
508 
509 	/* MCE errata HSD131, HSM142, HSW131, BDM48, HSM142 and SKX37 */
510 	if ((c->x86 == 6) &&
511 	    ((c->x86_model == INTEL_FAM6_HASWELL) ||
512 	     (c->x86_model == INTEL_FAM6_HASWELL_L) ||
513 	     (c->x86_model == INTEL_FAM6_BROADWELL) ||
514 	     (c->x86_model == INTEL_FAM6_HASWELL_G) ||
515 	     (c->x86_model == INTEL_FAM6_SKYLAKE_X)) &&
516 	    (m->bank == 0) &&
517 	    ((m->status & 0xa0000000ffffffff) == 0x80000000000f0005))
518 		return true;
519 
520 	return false;
521 }
522