1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* |
3 | * Machine check handler. |
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/thread_info.h> |
13 | #include <linux/capability.h> |
14 | #include <linux/miscdevice.h> |
15 | #include <linux/ratelimit.h> |
16 | #include <linux/rcupdate.h> |
17 | #include <linux/kobject.h> |
18 | #include <linux/uaccess.h> |
19 | #include <linux/kdebug.h> |
20 | #include <linux/kernel.h> |
21 | #include <linux/percpu.h> |
22 | #include <linux/string.h> |
23 | #include <linux/device.h> |
24 | #include <linux/syscore_ops.h> |
25 | #include <linux/delay.h> |
26 | #include <linux/ctype.h> |
27 | #include <linux/sched.h> |
28 | #include <linux/sysfs.h> |
29 | #include <linux/types.h> |
30 | #include <linux/slab.h> |
31 | #include <linux/init.h> |
32 | #include <linux/kmod.h> |
33 | #include <linux/poll.h> |
34 | #include <linux/nmi.h> |
35 | #include <linux/cpu.h> |
36 | #include <linux/ras.h> |
37 | #include <linux/smp.h> |
38 | #include <linux/fs.h> |
39 | #include <linux/mm.h> |
40 | #include <linux/debugfs.h> |
41 | #include <linux/irq_work.h> |
42 | #include <linux/export.h> |
43 | #include <linux/set_memory.h> |
44 | #include <linux/sync_core.h> |
45 | #include <linux/task_work.h> |
46 | #include <linux/hardirq.h> |
47 | |
48 | #include <asm/intel-family.h> |
49 | #include <asm/processor.h> |
50 | #include <asm/traps.h> |
51 | #include <asm/tlbflush.h> |
52 | #include <asm/mce.h> |
53 | #include <asm/msr.h> |
54 | #include <asm/reboot.h> |
55 | |
56 | #include "internal.h" |
57 | |
58 | /* sysfs synchronization */ |
59 | static DEFINE_MUTEX(mce_sysfs_mutex); |
60 | |
61 | #define CREATE_TRACE_POINTS |
62 | #include <trace/events/mce.h> |
63 | |
64 | #define SPINUNIT 100 /* 100ns */ |
65 | |
66 | DEFINE_PER_CPU(unsigned, mce_exception_count); |
67 | |
68 | DEFINE_PER_CPU_READ_MOSTLY(unsigned int, mce_num_banks); |
69 | |
70 | DEFINE_PER_CPU_READ_MOSTLY(struct mce_bank[MAX_NR_BANKS], mce_banks_array); |
71 | |
72 | #define ATTR_LEN 16 |
73 | /* One object for each MCE bank, shared by all CPUs */ |
74 | struct mce_bank_dev { |
75 | struct device_attribute attr; /* device attribute */ |
76 | char attrname[ATTR_LEN]; /* attribute name */ |
77 | u8 bank; /* bank number */ |
78 | }; |
79 | static struct mce_bank_dev mce_bank_devs[MAX_NR_BANKS]; |
80 | |
81 | struct mce_vendor_flags mce_flags __read_mostly; |
82 | |
83 | struct mca_config mca_cfg __read_mostly = { |
84 | .bootlog = -1, |
85 | .monarch_timeout = -1 |
86 | }; |
87 | |
88 | static DEFINE_PER_CPU(struct mce, mces_seen); |
89 | static unsigned long mce_need_notify; |
90 | |
91 | /* |
92 | * MCA banks polled by the period polling timer for corrected events. |
93 | * With Intel CMCI, this only has MCA banks which do not support CMCI (if any). |
94 | */ |
95 | DEFINE_PER_CPU(mce_banks_t, mce_poll_banks) = { |
96 | [0 ... BITS_TO_LONGS(MAX_NR_BANKS)-1] = ~0UL |
97 | }; |
98 | |
99 | /* |
100 | * MCA banks controlled through firmware first for corrected errors. |
101 | * This is a global list of banks for which we won't enable CMCI and we |
102 | * won't poll. Firmware controls these banks and is responsible for |
103 | * reporting corrected errors through GHES. Uncorrected/recoverable |
104 | * errors are still notified through a machine check. |
105 | */ |
106 | mce_banks_t mce_banks_ce_disabled; |
107 | |
108 | static struct work_struct mce_work; |
109 | static struct irq_work mce_irq_work; |
110 | |
111 | /* |
112 | * CPU/chipset specific EDAC code can register a notifier call here to print |
113 | * MCE errors in a human-readable form. |
114 | */ |
115 | BLOCKING_NOTIFIER_HEAD(x86_mce_decoder_chain); |
116 | |
117 | /* Do initial initialization of a struct mce */ |
118 | void mce_setup(struct mce *m) |
119 | { |
120 | memset(m, 0, sizeof(struct mce)); |
121 | m->cpu = m->extcpu = smp_processor_id(); |
122 | /* need the internal __ version to avoid deadlocks */ |
123 | m->time = __ktime_get_real_seconds(); |
124 | m->cpuvendor = boot_cpu_data.x86_vendor; |
125 | m->cpuid = cpuid_eax(op: 1); |
126 | m->socketid = cpu_data(m->extcpu).topo.pkg_id; |
127 | m->apicid = cpu_data(m->extcpu).topo.initial_apicid; |
128 | m->mcgcap = __rdmsr(MSR_IA32_MCG_CAP); |
129 | m->ppin = cpu_data(m->extcpu).ppin; |
130 | m->microcode = boot_cpu_data.microcode; |
131 | } |
132 | |
133 | DEFINE_PER_CPU(struct mce, injectm); |
134 | EXPORT_PER_CPU_SYMBOL_GPL(injectm); |
135 | |
136 | void mce_log(struct mce *m) |
137 | { |
138 | if (!mce_gen_pool_add(mce: m)) |
139 | irq_work_queue(work: &mce_irq_work); |
140 | } |
141 | EXPORT_SYMBOL_GPL(mce_log); |
142 | |
143 | void mce_register_decode_chain(struct notifier_block *nb) |
144 | { |
145 | if (WARN_ON(nb->priority < MCE_PRIO_LOWEST || |
146 | nb->priority > MCE_PRIO_HIGHEST)) |
147 | return; |
148 | |
149 | blocking_notifier_chain_register(nh: &x86_mce_decoder_chain, nb); |
150 | } |
151 | EXPORT_SYMBOL_GPL(mce_register_decode_chain); |
152 | |
153 | void mce_unregister_decode_chain(struct notifier_block *nb) |
154 | { |
155 | blocking_notifier_chain_unregister(nh: &x86_mce_decoder_chain, nb); |
156 | } |
157 | EXPORT_SYMBOL_GPL(mce_unregister_decode_chain); |
158 | |
159 | static void __print_mce(struct mce *m) |
160 | { |
161 | pr_emerg(HW_ERR "CPU %d: Machine Check%s: %Lx Bank %d: %016Lx\n" , |
162 | m->extcpu, |
163 | (m->mcgstatus & MCG_STATUS_MCIP ? " Exception" : "" ), |
164 | m->mcgstatus, m->bank, m->status); |
165 | |
166 | if (m->ip) { |
167 | pr_emerg(HW_ERR "RIP%s %02x:<%016Lx> " , |
168 | !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "" , |
169 | m->cs, m->ip); |
170 | |
171 | if (m->cs == __KERNEL_CS) |
172 | pr_cont("{%pS}" , (void *)(unsigned long)m->ip); |
173 | pr_cont("\n" ); |
174 | } |
175 | |
176 | pr_emerg(HW_ERR "TSC %llx " , m->tsc); |
177 | if (m->addr) |
178 | pr_cont("ADDR %llx " , m->addr); |
179 | if (m->misc) |
180 | pr_cont("MISC %llx " , m->misc); |
181 | if (m->ppin) |
182 | pr_cont("PPIN %llx " , m->ppin); |
183 | |
184 | if (mce_flags.smca) { |
185 | if (m->synd) |
186 | pr_cont("SYND %llx " , m->synd); |
187 | if (m->ipid) |
188 | pr_cont("IPID %llx " , m->ipid); |
189 | } |
190 | |
191 | pr_cont("\n" ); |
192 | |
193 | /* |
194 | * Note this output is parsed by external tools and old fields |
195 | * should not be changed. |
196 | */ |
197 | pr_emerg(HW_ERR "PROCESSOR %u:%x TIME %llu SOCKET %u APIC %x microcode %x\n" , |
198 | m->cpuvendor, m->cpuid, m->time, m->socketid, m->apicid, |
199 | m->microcode); |
200 | } |
201 | |
202 | static void print_mce(struct mce *m) |
203 | { |
204 | __print_mce(m); |
205 | |
206 | if (m->cpuvendor != X86_VENDOR_AMD && m->cpuvendor != X86_VENDOR_HYGON) |
207 | pr_emerg_ratelimited(HW_ERR "Run the above through 'mcelog --ascii'\n" ); |
208 | } |
209 | |
210 | #define PANIC_TIMEOUT 5 /* 5 seconds */ |
211 | |
212 | static atomic_t mce_panicked; |
213 | |
214 | static int fake_panic; |
215 | static atomic_t mce_fake_panicked; |
216 | |
217 | /* Panic in progress. Enable interrupts and wait for final IPI */ |
218 | static void wait_for_panic(void) |
219 | { |
220 | long timeout = PANIC_TIMEOUT*USEC_PER_SEC; |
221 | |
222 | preempt_disable(); |
223 | local_irq_enable(); |
224 | while (timeout-- > 0) |
225 | udelay(1); |
226 | if (panic_timeout == 0) |
227 | panic_timeout = mca_cfg.panic_timeout; |
228 | panic(fmt: "Panicing machine check CPU died" ); |
229 | } |
230 | |
231 | static noinstr void mce_panic(const char *msg, struct mce *final, char *exp) |
232 | { |
233 | struct llist_node *pending; |
234 | struct mce_evt_llist *l; |
235 | int apei_err = 0; |
236 | |
237 | /* |
238 | * Allow instrumentation around external facilities usage. Not that it |
239 | * matters a whole lot since the machine is going to panic anyway. |
240 | */ |
241 | instrumentation_begin(); |
242 | |
243 | if (!fake_panic) { |
244 | /* |
245 | * Make sure only one CPU runs in machine check panic |
246 | */ |
247 | if (atomic_inc_return(v: &mce_panicked) > 1) |
248 | wait_for_panic(); |
249 | barrier(); |
250 | |
251 | bust_spinlocks(yes: 1); |
252 | console_verbose(); |
253 | } else { |
254 | /* Don't log too much for fake panic */ |
255 | if (atomic_inc_return(v: &mce_fake_panicked) > 1) |
256 | goto out; |
257 | } |
258 | pending = mce_gen_pool_prepare_records(); |
259 | /* First print corrected ones that are still unlogged */ |
260 | llist_for_each_entry(l, pending, llnode) { |
261 | struct mce *m = &l->mce; |
262 | if (!(m->status & MCI_STATUS_UC)) { |
263 | print_mce(m); |
264 | if (!apei_err) |
265 | apei_err = apei_write_mce(m); |
266 | } |
267 | } |
268 | /* Now print uncorrected but with the final one last */ |
269 | llist_for_each_entry(l, pending, llnode) { |
270 | struct mce *m = &l->mce; |
271 | if (!(m->status & MCI_STATUS_UC)) |
272 | continue; |
273 | if (!final || mce_cmp(m1: m, m2: final)) { |
274 | print_mce(m); |
275 | if (!apei_err) |
276 | apei_err = apei_write_mce(m); |
277 | } |
278 | } |
279 | if (final) { |
280 | print_mce(m: final); |
281 | if (!apei_err) |
282 | apei_err = apei_write_mce(m: final); |
283 | } |
284 | if (exp) |
285 | pr_emerg(HW_ERR "Machine check: %s\n" , exp); |
286 | if (!fake_panic) { |
287 | if (panic_timeout == 0) |
288 | panic_timeout = mca_cfg.panic_timeout; |
289 | panic(fmt: msg); |
290 | } else |
291 | pr_emerg(HW_ERR "Fake kernel panic: %s\n" , msg); |
292 | |
293 | out: |
294 | instrumentation_end(); |
295 | } |
296 | |
297 | /* Support code for software error injection */ |
298 | |
299 | static int msr_to_offset(u32 msr) |
300 | { |
301 | unsigned bank = __this_cpu_read(injectm.bank); |
302 | |
303 | if (msr == mca_cfg.rip_msr) |
304 | return offsetof(struct mce, ip); |
305 | if (msr == mca_msr_reg(bank, reg: MCA_STATUS)) |
306 | return offsetof(struct mce, status); |
307 | if (msr == mca_msr_reg(bank, reg: MCA_ADDR)) |
308 | return offsetof(struct mce, addr); |
309 | if (msr == mca_msr_reg(bank, reg: MCA_MISC)) |
310 | return offsetof(struct mce, misc); |
311 | if (msr == MSR_IA32_MCG_STATUS) |
312 | return offsetof(struct mce, mcgstatus); |
313 | return -1; |
314 | } |
315 | |
316 | void ex_handler_msr_mce(struct pt_regs *regs, bool wrmsr) |
317 | { |
318 | if (wrmsr) { |
319 | pr_emerg("MSR access error: WRMSR to 0x%x (tried to write 0x%08x%08x) at rIP: 0x%lx (%pS)\n" , |
320 | (unsigned int)regs->cx, (unsigned int)regs->dx, (unsigned int)regs->ax, |
321 | regs->ip, (void *)regs->ip); |
322 | } else { |
323 | pr_emerg("MSR access error: RDMSR from 0x%x at rIP: 0x%lx (%pS)\n" , |
324 | (unsigned int)regs->cx, regs->ip, (void *)regs->ip); |
325 | } |
326 | |
327 | show_stack_regs(regs); |
328 | |
329 | panic(fmt: "MCA architectural violation!\n" ); |
330 | |
331 | while (true) |
332 | cpu_relax(); |
333 | } |
334 | |
335 | /* MSR access wrappers used for error injection */ |
336 | noinstr u64 mce_rdmsrl(u32 msr) |
337 | { |
338 | DECLARE_ARGS(val, low, high); |
339 | |
340 | if (__this_cpu_read(injectm.finished)) { |
341 | int offset; |
342 | u64 ret; |
343 | |
344 | instrumentation_begin(); |
345 | |
346 | offset = msr_to_offset(msr); |
347 | if (offset < 0) |
348 | ret = 0; |
349 | else |
350 | ret = *(u64 *)((char *)this_cpu_ptr(&injectm) + offset); |
351 | |
352 | instrumentation_end(); |
353 | |
354 | return ret; |
355 | } |
356 | |
357 | /* |
358 | * RDMSR on MCA MSRs should not fault. If they do, this is very much an |
359 | * architectural violation and needs to be reported to hw vendor. Panic |
360 | * the box to not allow any further progress. |
361 | */ |
362 | asm volatile("1: rdmsr\n" |
363 | "2:\n" |
364 | _ASM_EXTABLE_TYPE(1b, 2b, EX_TYPE_RDMSR_IN_MCE) |
365 | : EAX_EDX_RET(val, low, high) : "c" (msr)); |
366 | |
367 | |
368 | return EAX_EDX_VAL(val, low, high); |
369 | } |
370 | |
371 | static noinstr void mce_wrmsrl(u32 msr, u64 v) |
372 | { |
373 | u32 low, high; |
374 | |
375 | if (__this_cpu_read(injectm.finished)) { |
376 | int offset; |
377 | |
378 | instrumentation_begin(); |
379 | |
380 | offset = msr_to_offset(msr); |
381 | if (offset >= 0) |
382 | *(u64 *)((char *)this_cpu_ptr(&injectm) + offset) = v; |
383 | |
384 | instrumentation_end(); |
385 | |
386 | return; |
387 | } |
388 | |
389 | low = (u32)v; |
390 | high = (u32)(v >> 32); |
391 | |
392 | /* See comment in mce_rdmsrl() */ |
393 | asm volatile("1: wrmsr\n" |
394 | "2:\n" |
395 | _ASM_EXTABLE_TYPE(1b, 2b, EX_TYPE_WRMSR_IN_MCE) |
396 | : : "c" (msr), "a" (low), "d" (high) : "memory" ); |
397 | } |
398 | |
399 | /* |
400 | * Collect all global (w.r.t. this processor) status about this machine |
401 | * check into our "mce" struct so that we can use it later to assess |
402 | * the severity of the problem as we read per-bank specific details. |
403 | */ |
404 | static noinstr void mce_gather_info(struct mce *m, struct pt_regs *regs) |
405 | { |
406 | /* |
407 | * Enable instrumentation around mce_setup() which calls external |
408 | * facilities. |
409 | */ |
410 | instrumentation_begin(); |
411 | mce_setup(m); |
412 | instrumentation_end(); |
413 | |
414 | m->mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS); |
415 | if (regs) { |
416 | /* |
417 | * Get the address of the instruction at the time of |
418 | * the machine check error. |
419 | */ |
420 | if (m->mcgstatus & (MCG_STATUS_RIPV|MCG_STATUS_EIPV)) { |
421 | m->ip = regs->ip; |
422 | m->cs = regs->cs; |
423 | |
424 | /* |
425 | * When in VM86 mode make the cs look like ring 3 |
426 | * always. This is a lie, but it's better than passing |
427 | * the additional vm86 bit around everywhere. |
428 | */ |
429 | if (v8086_mode(regs)) |
430 | m->cs |= 3; |
431 | } |
432 | /* Use accurate RIP reporting if available. */ |
433 | if (mca_cfg.rip_msr) |
434 | m->ip = mce_rdmsrl(msr: mca_cfg.rip_msr); |
435 | } |
436 | } |
437 | |
438 | int mce_available(struct cpuinfo_x86 *c) |
439 | { |
440 | if (mca_cfg.disabled) |
441 | return 0; |
442 | return cpu_has(c, X86_FEATURE_MCE) && cpu_has(c, X86_FEATURE_MCA); |
443 | } |
444 | |
445 | static void mce_schedule_work(void) |
446 | { |
447 | if (!mce_gen_pool_empty()) |
448 | schedule_work(work: &mce_work); |
449 | } |
450 | |
451 | static void mce_irq_work_cb(struct irq_work *entry) |
452 | { |
453 | mce_schedule_work(); |
454 | } |
455 | |
456 | bool mce_usable_address(struct mce *m) |
457 | { |
458 | if (!(m->status & MCI_STATUS_ADDRV)) |
459 | return false; |
460 | |
461 | switch (m->cpuvendor) { |
462 | case X86_VENDOR_AMD: |
463 | return amd_mce_usable_address(m); |
464 | |
465 | case X86_VENDOR_INTEL: |
466 | case X86_VENDOR_ZHAOXIN: |
467 | return intel_mce_usable_address(m); |
468 | |
469 | default: |
470 | return true; |
471 | } |
472 | } |
473 | EXPORT_SYMBOL_GPL(mce_usable_address); |
474 | |
475 | bool mce_is_memory_error(struct mce *m) |
476 | { |
477 | switch (m->cpuvendor) { |
478 | case X86_VENDOR_AMD: |
479 | case X86_VENDOR_HYGON: |
480 | return amd_mce_is_memory_error(m); |
481 | |
482 | case X86_VENDOR_INTEL: |
483 | case X86_VENDOR_ZHAOXIN: |
484 | /* |
485 | * Intel SDM Volume 3B - 15.9.2 Compound Error Codes |
486 | * |
487 | * Bit 7 of the MCACOD field of IA32_MCi_STATUS is used for |
488 | * indicating a memory error. Bit 8 is used for indicating a |
489 | * cache hierarchy error. The combination of bit 2 and bit 3 |
490 | * is used for indicating a `generic' cache hierarchy error |
491 | * But we can't just blindly check the above bits, because if |
492 | * bit 11 is set, then it is a bus/interconnect error - and |
493 | * either way the above bits just gives more detail on what |
494 | * bus/interconnect error happened. Note that bit 12 can be |
495 | * ignored, as it's the "filter" bit. |
496 | */ |
497 | return (m->status & 0xef80) == BIT(7) || |
498 | (m->status & 0xef00) == BIT(8) || |
499 | (m->status & 0xeffc) == 0xc; |
500 | |
501 | default: |
502 | return false; |
503 | } |
504 | } |
505 | EXPORT_SYMBOL_GPL(mce_is_memory_error); |
506 | |
507 | static bool whole_page(struct mce *m) |
508 | { |
509 | if (!mca_cfg.ser || !(m->status & MCI_STATUS_MISCV)) |
510 | return true; |
511 | |
512 | return MCI_MISC_ADDR_LSB(m->misc) >= PAGE_SHIFT; |
513 | } |
514 | |
515 | bool mce_is_correctable(struct mce *m) |
516 | { |
517 | if (m->cpuvendor == X86_VENDOR_AMD && m->status & MCI_STATUS_DEFERRED) |
518 | return false; |
519 | |
520 | if (m->cpuvendor == X86_VENDOR_HYGON && m->status & MCI_STATUS_DEFERRED) |
521 | return false; |
522 | |
523 | if (m->status & MCI_STATUS_UC) |
524 | return false; |
525 | |
526 | return true; |
527 | } |
528 | EXPORT_SYMBOL_GPL(mce_is_correctable); |
529 | |
530 | static int mce_early_notifier(struct notifier_block *nb, unsigned long val, |
531 | void *data) |
532 | { |
533 | struct mce *m = (struct mce *)data; |
534 | |
535 | if (!m) |
536 | return NOTIFY_DONE; |
537 | |
538 | /* Emit the trace record: */ |
539 | trace_mce_record(m); |
540 | |
541 | set_bit(nr: 0, addr: &mce_need_notify); |
542 | |
543 | mce_notify_irq(); |
544 | |
545 | return NOTIFY_DONE; |
546 | } |
547 | |
548 | static struct notifier_block early_nb = { |
549 | .notifier_call = mce_early_notifier, |
550 | .priority = MCE_PRIO_EARLY, |
551 | }; |
552 | |
553 | static int uc_decode_notifier(struct notifier_block *nb, unsigned long val, |
554 | void *data) |
555 | { |
556 | struct mce *mce = (struct mce *)data; |
557 | unsigned long pfn; |
558 | |
559 | if (!mce || !mce_usable_address(mce)) |
560 | return NOTIFY_DONE; |
561 | |
562 | if (mce->severity != MCE_AO_SEVERITY && |
563 | mce->severity != MCE_DEFERRED_SEVERITY) |
564 | return NOTIFY_DONE; |
565 | |
566 | pfn = (mce->addr & MCI_ADDR_PHYSADDR) >> PAGE_SHIFT; |
567 | if (!memory_failure(pfn, flags: 0)) { |
568 | set_mce_nospec(pfn); |
569 | mce->kflags |= MCE_HANDLED_UC; |
570 | } |
571 | |
572 | return NOTIFY_OK; |
573 | } |
574 | |
575 | static struct notifier_block mce_uc_nb = { |
576 | .notifier_call = uc_decode_notifier, |
577 | .priority = MCE_PRIO_UC, |
578 | }; |
579 | |
580 | static int mce_default_notifier(struct notifier_block *nb, unsigned long val, |
581 | void *data) |
582 | { |
583 | struct mce *m = (struct mce *)data; |
584 | |
585 | if (!m) |
586 | return NOTIFY_DONE; |
587 | |
588 | if (mca_cfg.print_all || !m->kflags) |
589 | __print_mce(m); |
590 | |
591 | return NOTIFY_DONE; |
592 | } |
593 | |
594 | static struct notifier_block mce_default_nb = { |
595 | .notifier_call = mce_default_notifier, |
596 | /* lowest prio, we want it to run last. */ |
597 | .priority = MCE_PRIO_LOWEST, |
598 | }; |
599 | |
600 | /* |
601 | * Read ADDR and MISC registers. |
602 | */ |
603 | static noinstr void mce_read_aux(struct mce *m, int i) |
604 | { |
605 | if (m->status & MCI_STATUS_MISCV) |
606 | m->misc = mce_rdmsrl(msr: mca_msr_reg(bank: i, reg: MCA_MISC)); |
607 | |
608 | if (m->status & MCI_STATUS_ADDRV) { |
609 | m->addr = mce_rdmsrl(msr: mca_msr_reg(bank: i, reg: MCA_ADDR)); |
610 | |
611 | /* |
612 | * Mask the reported address by the reported granularity. |
613 | */ |
614 | if (mca_cfg.ser && (m->status & MCI_STATUS_MISCV)) { |
615 | u8 shift = MCI_MISC_ADDR_LSB(m->misc); |
616 | m->addr >>= shift; |
617 | m->addr <<= shift; |
618 | } |
619 | |
620 | smca_extract_err_addr(m); |
621 | } |
622 | |
623 | if (mce_flags.smca) { |
624 | m->ipid = mce_rdmsrl(MSR_AMD64_SMCA_MCx_IPID(i)); |
625 | |
626 | if (m->status & MCI_STATUS_SYNDV) |
627 | m->synd = mce_rdmsrl(MSR_AMD64_SMCA_MCx_SYND(i)); |
628 | } |
629 | } |
630 | |
631 | DEFINE_PER_CPU(unsigned, mce_poll_count); |
632 | |
633 | /* |
634 | * Poll for corrected events or events that happened before reset. |
635 | * Those are just logged through /dev/mcelog. |
636 | * |
637 | * This is executed in standard interrupt context. |
638 | * |
639 | * Note: spec recommends to panic for fatal unsignalled |
640 | * errors here. However this would be quite problematic -- |
641 | * we would need to reimplement the Monarch handling and |
642 | * it would mess up the exclusion between exception handler |
643 | * and poll handler -- * so we skip this for now. |
644 | * These cases should not happen anyways, or only when the CPU |
645 | * is already totally * confused. In this case it's likely it will |
646 | * not fully execute the machine check handler either. |
647 | */ |
648 | bool machine_check_poll(enum mcp_flags flags, mce_banks_t *b) |
649 | { |
650 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
651 | bool error_seen = false; |
652 | struct mce m; |
653 | int i; |
654 | |
655 | this_cpu_inc(mce_poll_count); |
656 | |
657 | mce_gather_info(m: &m, NULL); |
658 | |
659 | if (flags & MCP_TIMESTAMP) |
660 | m.tsc = rdtsc(); |
661 | |
662 | for (i = 0; i < this_cpu_read(mce_num_banks); i++) { |
663 | if (!mce_banks[i].ctl || !test_bit(i, *b)) |
664 | continue; |
665 | |
666 | m.misc = 0; |
667 | m.addr = 0; |
668 | m.bank = i; |
669 | |
670 | barrier(); |
671 | m.status = mce_rdmsrl(msr: mca_msr_reg(bank: i, reg: MCA_STATUS)); |
672 | |
673 | /* If this entry is not valid, ignore it */ |
674 | if (!(m.status & MCI_STATUS_VAL)) |
675 | continue; |
676 | |
677 | /* |
678 | * If we are logging everything (at CPU online) or this |
679 | * is a corrected error, then we must log it. |
680 | */ |
681 | if ((flags & MCP_UC) || !(m.status & MCI_STATUS_UC)) |
682 | goto log_it; |
683 | |
684 | /* |
685 | * Newer Intel systems that support software error |
686 | * recovery need to make additional checks. Other |
687 | * CPUs should skip over uncorrected errors, but log |
688 | * everything else. |
689 | */ |
690 | if (!mca_cfg.ser) { |
691 | if (m.status & MCI_STATUS_UC) |
692 | continue; |
693 | goto log_it; |
694 | } |
695 | |
696 | /* Log "not enabled" (speculative) errors */ |
697 | if (!(m.status & MCI_STATUS_EN)) |
698 | goto log_it; |
699 | |
700 | /* |
701 | * Log UCNA (SDM: 15.6.3 "UCR Error Classification") |
702 | * UC == 1 && PCC == 0 && S == 0 |
703 | */ |
704 | if (!(m.status & MCI_STATUS_PCC) && !(m.status & MCI_STATUS_S)) |
705 | goto log_it; |
706 | |
707 | /* |
708 | * Skip anything else. Presumption is that our read of this |
709 | * bank is racing with a machine check. Leave the log alone |
710 | * for do_machine_check() to deal with it. |
711 | */ |
712 | continue; |
713 | |
714 | log_it: |
715 | error_seen = true; |
716 | |
717 | if (flags & MCP_DONTLOG) |
718 | goto clear_it; |
719 | |
720 | mce_read_aux(m: &m, i); |
721 | m.severity = mce_severity(a: &m, NULL, NULL, is_excp: false); |
722 | /* |
723 | * Don't get the IP here because it's unlikely to |
724 | * have anything to do with the actual error location. |
725 | */ |
726 | |
727 | if (mca_cfg.dont_log_ce && !mce_usable_address(&m)) |
728 | goto clear_it; |
729 | |
730 | if (flags & MCP_QUEUE_LOG) |
731 | mce_gen_pool_add(mce: &m); |
732 | else |
733 | mce_log(&m); |
734 | |
735 | clear_it: |
736 | /* |
737 | * Clear state for this bank. |
738 | */ |
739 | mce_wrmsrl(msr: mca_msr_reg(bank: i, reg: MCA_STATUS), v: 0); |
740 | } |
741 | |
742 | /* |
743 | * Don't clear MCG_STATUS here because it's only defined for |
744 | * exceptions. |
745 | */ |
746 | |
747 | sync_core(); |
748 | |
749 | return error_seen; |
750 | } |
751 | EXPORT_SYMBOL_GPL(machine_check_poll); |
752 | |
753 | /* |
754 | * During IFU recovery Sandy Bridge -EP4S processors set the RIPV and |
755 | * EIPV bits in MCG_STATUS to zero on the affected logical processor (SDM |
756 | * Vol 3B Table 15-20). But this confuses both the code that determines |
757 | * whether the machine check occurred in kernel or user mode, and also |
758 | * the severity assessment code. Pretend that EIPV was set, and take the |
759 | * ip/cs values from the pt_regs that mce_gather_info() ignored earlier. |
760 | */ |
761 | static __always_inline void |
762 | quirk_sandybridge_ifu(int bank, struct mce *m, struct pt_regs *regs) |
763 | { |
764 | if (bank != 0) |
765 | return; |
766 | if ((m->mcgstatus & (MCG_STATUS_EIPV|MCG_STATUS_RIPV)) != 0) |
767 | return; |
768 | if ((m->status & (MCI_STATUS_OVER|MCI_STATUS_UC| |
769 | MCI_STATUS_EN|MCI_STATUS_MISCV|MCI_STATUS_ADDRV| |
770 | MCI_STATUS_PCC|MCI_STATUS_S|MCI_STATUS_AR| |
771 | MCACOD)) != |
772 | (MCI_STATUS_UC|MCI_STATUS_EN| |
773 | MCI_STATUS_MISCV|MCI_STATUS_ADDRV|MCI_STATUS_S| |
774 | MCI_STATUS_AR|MCACOD_INSTR)) |
775 | return; |
776 | |
777 | m->mcgstatus |= MCG_STATUS_EIPV; |
778 | m->ip = regs->ip; |
779 | m->cs = regs->cs; |
780 | } |
781 | |
782 | /* |
783 | * Disable fast string copy and return from the MCE handler upon the first SRAR |
784 | * MCE on bank 1 due to a CPU erratum on Intel Skylake/Cascade Lake/Cooper Lake |
785 | * CPUs. |
786 | * The fast string copy instructions ("REP; MOVS*") could consume an |
787 | * uncorrectable memory error in the cache line _right after_ the desired region |
788 | * to copy and raise an MCE with RIP pointing to the instruction _after_ the |
789 | * "REP; MOVS*". |
790 | * This mitigation addresses the issue completely with the caveat of performance |
791 | * degradation on the CPU affected. This is still better than the OS crashing on |
792 | * MCEs raised on an irrelevant process due to "REP; MOVS*" accesses from a |
793 | * kernel context (e.g., copy_page). |
794 | * |
795 | * Returns true when fast string copy on CPU has been disabled. |
796 | */ |
797 | static noinstr bool quirk_skylake_repmov(void) |
798 | { |
799 | u64 mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS); |
800 | u64 misc_enable = mce_rdmsrl(MSR_IA32_MISC_ENABLE); |
801 | u64 mc1_status; |
802 | |
803 | /* |
804 | * Apply the quirk only to local machine checks, i.e., no broadcast |
805 | * sync is needed. |
806 | */ |
807 | if (!(mcgstatus & MCG_STATUS_LMCES) || |
808 | !(misc_enable & MSR_IA32_MISC_ENABLE_FAST_STRING)) |
809 | return false; |
810 | |
811 | mc1_status = mce_rdmsrl(MSR_IA32_MCx_STATUS(1)); |
812 | |
813 | /* Check for a software-recoverable data fetch error. */ |
814 | if ((mc1_status & |
815 | (MCI_STATUS_VAL | MCI_STATUS_OVER | MCI_STATUS_UC | MCI_STATUS_EN | |
816 | MCI_STATUS_ADDRV | MCI_STATUS_MISCV | MCI_STATUS_PCC | |
817 | MCI_STATUS_AR | MCI_STATUS_S)) == |
818 | (MCI_STATUS_VAL | MCI_STATUS_UC | MCI_STATUS_EN | |
819 | MCI_STATUS_ADDRV | MCI_STATUS_MISCV | |
820 | MCI_STATUS_AR | MCI_STATUS_S)) { |
821 | misc_enable &= ~MSR_IA32_MISC_ENABLE_FAST_STRING; |
822 | mce_wrmsrl(MSR_IA32_MISC_ENABLE, v: misc_enable); |
823 | mce_wrmsrl(MSR_IA32_MCx_STATUS(1), v: 0); |
824 | |
825 | instrumentation_begin(); |
826 | pr_err_once("Erratum detected, disable fast string copy instructions.\n" ); |
827 | instrumentation_end(); |
828 | |
829 | return true; |
830 | } |
831 | |
832 | return false; |
833 | } |
834 | |
835 | /* |
836 | * Some Zen-based Instruction Fetch Units set EIPV=RIPV=0 on poison consumption |
837 | * errors. This means mce_gather_info() will not save the "ip" and "cs" registers. |
838 | * |
839 | * However, the context is still valid, so save the "cs" register for later use. |
840 | * |
841 | * The "ip" register is truly unknown, so don't save it or fixup EIPV/RIPV. |
842 | * |
843 | * The Instruction Fetch Unit is at MCA bank 1 for all affected systems. |
844 | */ |
845 | static __always_inline void quirk_zen_ifu(int bank, struct mce *m, struct pt_regs *regs) |
846 | { |
847 | if (bank != 1) |
848 | return; |
849 | if (!(m->status & MCI_STATUS_POISON)) |
850 | return; |
851 | |
852 | m->cs = regs->cs; |
853 | } |
854 | |
855 | /* |
856 | * Do a quick check if any of the events requires a panic. |
857 | * This decides if we keep the events around or clear them. |
858 | */ |
859 | static __always_inline int mce_no_way_out(struct mce *m, char **msg, unsigned long *validp, |
860 | struct pt_regs *regs) |
861 | { |
862 | char *tmp = *msg; |
863 | int i; |
864 | |
865 | for (i = 0; i < this_cpu_read(mce_num_banks); i++) { |
866 | m->status = mce_rdmsrl(msr: mca_msr_reg(bank: i, reg: MCA_STATUS)); |
867 | if (!(m->status & MCI_STATUS_VAL)) |
868 | continue; |
869 | |
870 | arch___set_bit(nr: i, addr: validp); |
871 | if (mce_flags.snb_ifu_quirk) |
872 | quirk_sandybridge_ifu(bank: i, m, regs); |
873 | |
874 | if (mce_flags.zen_ifu_quirk) |
875 | quirk_zen_ifu(bank: i, m, regs); |
876 | |
877 | m->bank = i; |
878 | if (mce_severity(a: m, regs, msg: &tmp, is_excp: true) >= MCE_PANIC_SEVERITY) { |
879 | mce_read_aux(m, i); |
880 | *msg = tmp; |
881 | return 1; |
882 | } |
883 | } |
884 | return 0; |
885 | } |
886 | |
887 | /* |
888 | * Variable to establish order between CPUs while scanning. |
889 | * Each CPU spins initially until executing is equal its number. |
890 | */ |
891 | static atomic_t mce_executing; |
892 | |
893 | /* |
894 | * Defines order of CPUs on entry. First CPU becomes Monarch. |
895 | */ |
896 | static atomic_t mce_callin; |
897 | |
898 | /* |
899 | * Track which CPUs entered the MCA broadcast synchronization and which not in |
900 | * order to print holdouts. |
901 | */ |
902 | static cpumask_t mce_missing_cpus = CPU_MASK_ALL; |
903 | |
904 | /* |
905 | * Check if a timeout waiting for other CPUs happened. |
906 | */ |
907 | static noinstr int mce_timed_out(u64 *t, const char *msg) |
908 | { |
909 | int ret = 0; |
910 | |
911 | /* Enable instrumentation around calls to external facilities */ |
912 | instrumentation_begin(); |
913 | |
914 | /* |
915 | * The others already did panic for some reason. |
916 | * Bail out like in a timeout. |
917 | * rmb() to tell the compiler that system_state |
918 | * might have been modified by someone else. |
919 | */ |
920 | rmb(); |
921 | if (atomic_read(v: &mce_panicked)) |
922 | wait_for_panic(); |
923 | if (!mca_cfg.monarch_timeout) |
924 | goto out; |
925 | if ((s64)*t < SPINUNIT) { |
926 | if (cpumask_and(dstp: &mce_missing_cpus, cpu_online_mask, src2p: &mce_missing_cpus)) |
927 | pr_emerg("CPUs not responding to MCE broadcast (may include false positives): %*pbl\n" , |
928 | cpumask_pr_args(&mce_missing_cpus)); |
929 | mce_panic(msg, NULL, NULL); |
930 | |
931 | ret = 1; |
932 | goto out; |
933 | } |
934 | *t -= SPINUNIT; |
935 | |
936 | out: |
937 | touch_nmi_watchdog(); |
938 | |
939 | instrumentation_end(); |
940 | |
941 | return ret; |
942 | } |
943 | |
944 | /* |
945 | * The Monarch's reign. The Monarch is the CPU who entered |
946 | * the machine check handler first. It waits for the others to |
947 | * raise the exception too and then grades them. When any |
948 | * error is fatal panic. Only then let the others continue. |
949 | * |
950 | * The other CPUs entering the MCE handler will be controlled by the |
951 | * Monarch. They are called Subjects. |
952 | * |
953 | * This way we prevent any potential data corruption in a unrecoverable case |
954 | * and also makes sure always all CPU's errors are examined. |
955 | * |
956 | * Also this detects the case of a machine check event coming from outer |
957 | * space (not detected by any CPUs) In this case some external agent wants |
958 | * us to shut down, so panic too. |
959 | * |
960 | * The other CPUs might still decide to panic if the handler happens |
961 | * in a unrecoverable place, but in this case the system is in a semi-stable |
962 | * state and won't corrupt anything by itself. It's ok to let the others |
963 | * continue for a bit first. |
964 | * |
965 | * All the spin loops have timeouts; when a timeout happens a CPU |
966 | * typically elects itself to be Monarch. |
967 | */ |
968 | static void mce_reign(void) |
969 | { |
970 | int cpu; |
971 | struct mce *m = NULL; |
972 | int global_worst = 0; |
973 | char *msg = NULL; |
974 | |
975 | /* |
976 | * This CPU is the Monarch and the other CPUs have run |
977 | * through their handlers. |
978 | * Grade the severity of the errors of all the CPUs. |
979 | */ |
980 | for_each_possible_cpu(cpu) { |
981 | struct mce *mtmp = &per_cpu(mces_seen, cpu); |
982 | |
983 | if (mtmp->severity > global_worst) { |
984 | global_worst = mtmp->severity; |
985 | m = &per_cpu(mces_seen, cpu); |
986 | } |
987 | } |
988 | |
989 | /* |
990 | * Cannot recover? Panic here then. |
991 | * This dumps all the mces in the log buffer and stops the |
992 | * other CPUs. |
993 | */ |
994 | if (m && global_worst >= MCE_PANIC_SEVERITY) { |
995 | /* call mce_severity() to get "msg" for panic */ |
996 | mce_severity(a: m, NULL, msg: &msg, is_excp: true); |
997 | mce_panic(msg: "Fatal machine check" , final: m, exp: msg); |
998 | } |
999 | |
1000 | /* |
1001 | * For UC somewhere we let the CPU who detects it handle it. |
1002 | * Also must let continue the others, otherwise the handling |
1003 | * CPU could deadlock on a lock. |
1004 | */ |
1005 | |
1006 | /* |
1007 | * No machine check event found. Must be some external |
1008 | * source or one CPU is hung. Panic. |
1009 | */ |
1010 | if (global_worst <= MCE_KEEP_SEVERITY) |
1011 | mce_panic(msg: "Fatal machine check from unknown source" , NULL, NULL); |
1012 | |
1013 | /* |
1014 | * Now clear all the mces_seen so that they don't reappear on |
1015 | * the next mce. |
1016 | */ |
1017 | for_each_possible_cpu(cpu) |
1018 | memset(&per_cpu(mces_seen, cpu), 0, sizeof(struct mce)); |
1019 | } |
1020 | |
1021 | static atomic_t global_nwo; |
1022 | |
1023 | /* |
1024 | * Start of Monarch synchronization. This waits until all CPUs have |
1025 | * entered the exception handler and then determines if any of them |
1026 | * saw a fatal event that requires panic. Then it executes them |
1027 | * in the entry order. |
1028 | * TBD double check parallel CPU hotunplug |
1029 | */ |
1030 | static noinstr int mce_start(int *no_way_out) |
1031 | { |
1032 | u64 timeout = (u64)mca_cfg.monarch_timeout * NSEC_PER_USEC; |
1033 | int order, ret = -1; |
1034 | |
1035 | if (!timeout) |
1036 | return ret; |
1037 | |
1038 | raw_atomic_add(i: *no_way_out, v: &global_nwo); |
1039 | /* |
1040 | * Rely on the implied barrier below, such that global_nwo |
1041 | * is updated before mce_callin. |
1042 | */ |
1043 | order = raw_atomic_inc_return(v: &mce_callin); |
1044 | arch_cpumask_clear_cpu(smp_processor_id(), dstp: &mce_missing_cpus); |
1045 | |
1046 | /* Enable instrumentation around calls to external facilities */ |
1047 | instrumentation_begin(); |
1048 | |
1049 | /* |
1050 | * Wait for everyone. |
1051 | */ |
1052 | while (raw_atomic_read(v: &mce_callin) != num_online_cpus()) { |
1053 | if (mce_timed_out(t: &timeout, |
1054 | msg: "Timeout: Not all CPUs entered broadcast exception handler" )) { |
1055 | raw_atomic_set(v: &global_nwo, i: 0); |
1056 | goto out; |
1057 | } |
1058 | ndelay(SPINUNIT); |
1059 | } |
1060 | |
1061 | /* |
1062 | * mce_callin should be read before global_nwo |
1063 | */ |
1064 | smp_rmb(); |
1065 | |
1066 | if (order == 1) { |
1067 | /* |
1068 | * Monarch: Starts executing now, the others wait. |
1069 | */ |
1070 | raw_atomic_set(v: &mce_executing, i: 1); |
1071 | } else { |
1072 | /* |
1073 | * Subject: Now start the scanning loop one by one in |
1074 | * the original callin order. |
1075 | * This way when there are any shared banks it will be |
1076 | * only seen by one CPU before cleared, avoiding duplicates. |
1077 | */ |
1078 | while (raw_atomic_read(v: &mce_executing) < order) { |
1079 | if (mce_timed_out(t: &timeout, |
1080 | msg: "Timeout: Subject CPUs unable to finish machine check processing" )) { |
1081 | raw_atomic_set(v: &global_nwo, i: 0); |
1082 | goto out; |
1083 | } |
1084 | ndelay(SPINUNIT); |
1085 | } |
1086 | } |
1087 | |
1088 | /* |
1089 | * Cache the global no_way_out state. |
1090 | */ |
1091 | *no_way_out = raw_atomic_read(v: &global_nwo); |
1092 | |
1093 | ret = order; |
1094 | |
1095 | out: |
1096 | instrumentation_end(); |
1097 | |
1098 | return ret; |
1099 | } |
1100 | |
1101 | /* |
1102 | * Synchronize between CPUs after main scanning loop. |
1103 | * This invokes the bulk of the Monarch processing. |
1104 | */ |
1105 | static noinstr int mce_end(int order) |
1106 | { |
1107 | u64 timeout = (u64)mca_cfg.monarch_timeout * NSEC_PER_USEC; |
1108 | int ret = -1; |
1109 | |
1110 | /* Allow instrumentation around external facilities. */ |
1111 | instrumentation_begin(); |
1112 | |
1113 | if (!timeout) |
1114 | goto reset; |
1115 | if (order < 0) |
1116 | goto reset; |
1117 | |
1118 | /* |
1119 | * Allow others to run. |
1120 | */ |
1121 | atomic_inc(v: &mce_executing); |
1122 | |
1123 | if (order == 1) { |
1124 | /* |
1125 | * Monarch: Wait for everyone to go through their scanning |
1126 | * loops. |
1127 | */ |
1128 | while (atomic_read(v: &mce_executing) <= num_online_cpus()) { |
1129 | if (mce_timed_out(t: &timeout, |
1130 | msg: "Timeout: Monarch CPU unable to finish machine check processing" )) |
1131 | goto reset; |
1132 | ndelay(SPINUNIT); |
1133 | } |
1134 | |
1135 | mce_reign(); |
1136 | barrier(); |
1137 | ret = 0; |
1138 | } else { |
1139 | /* |
1140 | * Subject: Wait for Monarch to finish. |
1141 | */ |
1142 | while (atomic_read(v: &mce_executing) != 0) { |
1143 | if (mce_timed_out(t: &timeout, |
1144 | msg: "Timeout: Monarch CPU did not finish machine check processing" )) |
1145 | goto reset; |
1146 | ndelay(SPINUNIT); |
1147 | } |
1148 | |
1149 | /* |
1150 | * Don't reset anything. That's done by the Monarch. |
1151 | */ |
1152 | ret = 0; |
1153 | goto out; |
1154 | } |
1155 | |
1156 | /* |
1157 | * Reset all global state. |
1158 | */ |
1159 | reset: |
1160 | atomic_set(v: &global_nwo, i: 0); |
1161 | atomic_set(v: &mce_callin, i: 0); |
1162 | cpumask_setall(dstp: &mce_missing_cpus); |
1163 | barrier(); |
1164 | |
1165 | /* |
1166 | * Let others run again. |
1167 | */ |
1168 | atomic_set(v: &mce_executing, i: 0); |
1169 | |
1170 | out: |
1171 | instrumentation_end(); |
1172 | |
1173 | return ret; |
1174 | } |
1175 | |
1176 | static __always_inline void mce_clear_state(unsigned long *toclear) |
1177 | { |
1178 | int i; |
1179 | |
1180 | for (i = 0; i < this_cpu_read(mce_num_banks); i++) { |
1181 | if (arch_test_bit(nr: i, addr: toclear)) |
1182 | mce_wrmsrl(msr: mca_msr_reg(bank: i, reg: MCA_STATUS), v: 0); |
1183 | } |
1184 | } |
1185 | |
1186 | /* |
1187 | * Cases where we avoid rendezvous handler timeout: |
1188 | * 1) If this CPU is offline. |
1189 | * |
1190 | * 2) If crashing_cpu was set, e.g. we're entering kdump and we need to |
1191 | * skip those CPUs which remain looping in the 1st kernel - see |
1192 | * crash_nmi_callback(). |
1193 | * |
1194 | * Note: there still is a small window between kexec-ing and the new, |
1195 | * kdump kernel establishing a new #MC handler where a broadcasted MCE |
1196 | * might not get handled properly. |
1197 | */ |
1198 | static noinstr bool mce_check_crashing_cpu(void) |
1199 | { |
1200 | unsigned int cpu = smp_processor_id(); |
1201 | |
1202 | if (arch_cpu_is_offline(cpu) || |
1203 | (crashing_cpu != -1 && crashing_cpu != cpu)) { |
1204 | u64 mcgstatus; |
1205 | |
1206 | mcgstatus = __rdmsr(MSR_IA32_MCG_STATUS); |
1207 | |
1208 | if (boot_cpu_data.x86_vendor == X86_VENDOR_ZHAOXIN) { |
1209 | if (mcgstatus & MCG_STATUS_LMCES) |
1210 | return false; |
1211 | } |
1212 | |
1213 | if (mcgstatus & MCG_STATUS_RIPV) { |
1214 | __wrmsr(MSR_IA32_MCG_STATUS, low: 0, high: 0); |
1215 | return true; |
1216 | } |
1217 | } |
1218 | return false; |
1219 | } |
1220 | |
1221 | static __always_inline int |
1222 | __mc_scan_banks(struct mce *m, struct pt_regs *regs, struct mce *final, |
1223 | unsigned long *toclear, unsigned long *valid_banks, int no_way_out, |
1224 | int *worst) |
1225 | { |
1226 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
1227 | struct mca_config *cfg = &mca_cfg; |
1228 | int severity, i, taint = 0; |
1229 | |
1230 | for (i = 0; i < this_cpu_read(mce_num_banks); i++) { |
1231 | arch___clear_bit(nr: i, addr: toclear); |
1232 | if (!arch_test_bit(nr: i, addr: valid_banks)) |
1233 | continue; |
1234 | |
1235 | if (!mce_banks[i].ctl) |
1236 | continue; |
1237 | |
1238 | m->misc = 0; |
1239 | m->addr = 0; |
1240 | m->bank = i; |
1241 | |
1242 | m->status = mce_rdmsrl(msr: mca_msr_reg(bank: i, reg: MCA_STATUS)); |
1243 | if (!(m->status & MCI_STATUS_VAL)) |
1244 | continue; |
1245 | |
1246 | /* |
1247 | * Corrected or non-signaled errors are handled by |
1248 | * machine_check_poll(). Leave them alone, unless this panics. |
1249 | */ |
1250 | if (!(m->status & (cfg->ser ? MCI_STATUS_S : MCI_STATUS_UC)) && |
1251 | !no_way_out) |
1252 | continue; |
1253 | |
1254 | /* Set taint even when machine check was not enabled. */ |
1255 | taint++; |
1256 | |
1257 | severity = mce_severity(a: m, regs, NULL, is_excp: true); |
1258 | |
1259 | /* |
1260 | * When machine check was for corrected/deferred handler don't |
1261 | * touch, unless we're panicking. |
1262 | */ |
1263 | if ((severity == MCE_KEEP_SEVERITY || |
1264 | severity == MCE_UCNA_SEVERITY) && !no_way_out) |
1265 | continue; |
1266 | |
1267 | arch___set_bit(nr: i, addr: toclear); |
1268 | |
1269 | /* Machine check event was not enabled. Clear, but ignore. */ |
1270 | if (severity == MCE_NO_SEVERITY) |
1271 | continue; |
1272 | |
1273 | mce_read_aux(m, i); |
1274 | |
1275 | /* assuming valid severity level != 0 */ |
1276 | m->severity = severity; |
1277 | |
1278 | /* |
1279 | * Enable instrumentation around the mce_log() call which is |
1280 | * done in #MC context, where instrumentation is disabled. |
1281 | */ |
1282 | instrumentation_begin(); |
1283 | mce_log(m); |
1284 | instrumentation_end(); |
1285 | |
1286 | if (severity > *worst) { |
1287 | *final = *m; |
1288 | *worst = severity; |
1289 | } |
1290 | } |
1291 | |
1292 | /* mce_clear_state will clear *final, save locally for use later */ |
1293 | *m = *final; |
1294 | |
1295 | return taint; |
1296 | } |
1297 | |
1298 | static void kill_me_now(struct callback_head *ch) |
1299 | { |
1300 | struct task_struct *p = container_of(ch, struct task_struct, mce_kill_me); |
1301 | |
1302 | p->mce_count = 0; |
1303 | force_sig(SIGBUS); |
1304 | } |
1305 | |
1306 | static void kill_me_maybe(struct callback_head *cb) |
1307 | { |
1308 | struct task_struct *p = container_of(cb, struct task_struct, mce_kill_me); |
1309 | int flags = MF_ACTION_REQUIRED; |
1310 | unsigned long pfn; |
1311 | int ret; |
1312 | |
1313 | p->mce_count = 0; |
1314 | pr_err("Uncorrected hardware memory error in user-access at %llx" , p->mce_addr); |
1315 | |
1316 | if (!p->mce_ripv) |
1317 | flags |= MF_MUST_KILL; |
1318 | |
1319 | pfn = (p->mce_addr & MCI_ADDR_PHYSADDR) >> PAGE_SHIFT; |
1320 | ret = memory_failure(pfn, flags); |
1321 | if (!ret) { |
1322 | set_mce_nospec(pfn); |
1323 | sync_core(); |
1324 | return; |
1325 | } |
1326 | |
1327 | /* |
1328 | * -EHWPOISON from memory_failure() means that it already sent SIGBUS |
1329 | * to the current process with the proper error info, |
1330 | * -EOPNOTSUPP means hwpoison_filter() filtered the error event, |
1331 | * |
1332 | * In both cases, no further processing is required. |
1333 | */ |
1334 | if (ret == -EHWPOISON || ret == -EOPNOTSUPP) |
1335 | return; |
1336 | |
1337 | pr_err("Memory error not recovered" ); |
1338 | kill_me_now(ch: cb); |
1339 | } |
1340 | |
1341 | static void kill_me_never(struct callback_head *cb) |
1342 | { |
1343 | struct task_struct *p = container_of(cb, struct task_struct, mce_kill_me); |
1344 | unsigned long pfn; |
1345 | |
1346 | p->mce_count = 0; |
1347 | pr_err("Kernel accessed poison in user space at %llx\n" , p->mce_addr); |
1348 | pfn = (p->mce_addr & MCI_ADDR_PHYSADDR) >> PAGE_SHIFT; |
1349 | if (!memory_failure(pfn, flags: 0)) |
1350 | set_mce_nospec(pfn); |
1351 | } |
1352 | |
1353 | static void queue_task_work(struct mce *m, char *msg, void (*func)(struct callback_head *)) |
1354 | { |
1355 | int count = ++current->mce_count; |
1356 | |
1357 | /* First call, save all the details */ |
1358 | if (count == 1) { |
1359 | current->mce_addr = m->addr; |
1360 | current->mce_kflags = m->kflags; |
1361 | current->mce_ripv = !!(m->mcgstatus & MCG_STATUS_RIPV); |
1362 | current->mce_whole_page = whole_page(m); |
1363 | current->mce_kill_me.func = func; |
1364 | } |
1365 | |
1366 | /* Ten is likely overkill. Don't expect more than two faults before task_work() */ |
1367 | if (count > 10) |
1368 | mce_panic(msg: "Too many consecutive machine checks while accessing user data" , final: m, exp: msg); |
1369 | |
1370 | /* Second or later call, make sure page address matches the one from first call */ |
1371 | if (count > 1 && (current->mce_addr >> PAGE_SHIFT) != (m->addr >> PAGE_SHIFT)) |
1372 | mce_panic(msg: "Consecutive machine checks to different user pages" , final: m, exp: msg); |
1373 | |
1374 | /* Do not call task_work_add() more than once */ |
1375 | if (count > 1) |
1376 | return; |
1377 | |
1378 | task_work_add(current, twork: ¤t->mce_kill_me, mode: TWA_RESUME); |
1379 | } |
1380 | |
1381 | /* Handle unconfigured int18 (should never happen) */ |
1382 | static noinstr void unexpected_machine_check(struct pt_regs *regs) |
1383 | { |
1384 | instrumentation_begin(); |
1385 | pr_err("CPU#%d: Unexpected int18 (Machine Check)\n" , |
1386 | smp_processor_id()); |
1387 | instrumentation_end(); |
1388 | } |
1389 | |
1390 | /* |
1391 | * The actual machine check handler. This only handles real exceptions when |
1392 | * something got corrupted coming in through int 18. |
1393 | * |
1394 | * This is executed in #MC context not subject to normal locking rules. |
1395 | * This implies that most kernel services cannot be safely used. Don't even |
1396 | * think about putting a printk in there! |
1397 | * |
1398 | * On Intel systems this is entered on all CPUs in parallel through |
1399 | * MCE broadcast. However some CPUs might be broken beyond repair, |
1400 | * so be always careful when synchronizing with others. |
1401 | * |
1402 | * Tracing and kprobes are disabled: if we interrupted a kernel context |
1403 | * with IF=1, we need to minimize stack usage. There are also recursion |
1404 | * issues: if the machine check was due to a failure of the memory |
1405 | * backing the user stack, tracing that reads the user stack will cause |
1406 | * potentially infinite recursion. |
1407 | * |
1408 | * Currently, the #MC handler calls out to a number of external facilities |
1409 | * and, therefore, allows instrumentation around them. The optimal thing to |
1410 | * have would be to do the absolutely minimal work required in #MC context |
1411 | * and have instrumentation disabled only around that. Further processing can |
1412 | * then happen in process context where instrumentation is allowed. Achieving |
1413 | * that requires careful auditing and modifications. Until then, the code |
1414 | * allows instrumentation temporarily, where required. * |
1415 | */ |
1416 | noinstr void do_machine_check(struct pt_regs *regs) |
1417 | { |
1418 | int worst = 0, order, no_way_out, kill_current_task, lmce, taint = 0; |
1419 | DECLARE_BITMAP(valid_banks, MAX_NR_BANKS) = { 0 }; |
1420 | DECLARE_BITMAP(toclear, MAX_NR_BANKS) = { 0 }; |
1421 | struct mce m, *final; |
1422 | char *msg = NULL; |
1423 | |
1424 | if (unlikely(mce_flags.p5)) |
1425 | return pentium_machine_check(regs); |
1426 | else if (unlikely(mce_flags.winchip)) |
1427 | return winchip_machine_check(regs); |
1428 | else if (unlikely(!mca_cfg.initialized)) |
1429 | return unexpected_machine_check(regs); |
1430 | |
1431 | if (mce_flags.skx_repmov_quirk && quirk_skylake_repmov()) |
1432 | goto clear; |
1433 | |
1434 | /* |
1435 | * Establish sequential order between the CPUs entering the machine |
1436 | * check handler. |
1437 | */ |
1438 | order = -1; |
1439 | |
1440 | /* |
1441 | * If no_way_out gets set, there is no safe way to recover from this |
1442 | * MCE. |
1443 | */ |
1444 | no_way_out = 0; |
1445 | |
1446 | /* |
1447 | * If kill_current_task is not set, there might be a way to recover from this |
1448 | * error. |
1449 | */ |
1450 | kill_current_task = 0; |
1451 | |
1452 | /* |
1453 | * MCEs are always local on AMD. Same is determined by MCG_STATUS_LMCES |
1454 | * on Intel. |
1455 | */ |
1456 | lmce = 1; |
1457 | |
1458 | this_cpu_inc(mce_exception_count); |
1459 | |
1460 | mce_gather_info(m: &m, regs); |
1461 | m.tsc = rdtsc(); |
1462 | |
1463 | final = this_cpu_ptr(&mces_seen); |
1464 | *final = m; |
1465 | |
1466 | no_way_out = mce_no_way_out(m: &m, msg: &msg, validp: valid_banks, regs); |
1467 | |
1468 | barrier(); |
1469 | |
1470 | /* |
1471 | * When no restart IP might need to kill or panic. |
1472 | * Assume the worst for now, but if we find the |
1473 | * severity is MCE_AR_SEVERITY we have other options. |
1474 | */ |
1475 | if (!(m.mcgstatus & MCG_STATUS_RIPV)) |
1476 | kill_current_task = 1; |
1477 | /* |
1478 | * Check if this MCE is signaled to only this logical processor, |
1479 | * on Intel, Zhaoxin only. |
1480 | */ |
1481 | if (m.cpuvendor == X86_VENDOR_INTEL || |
1482 | m.cpuvendor == X86_VENDOR_ZHAOXIN) |
1483 | lmce = m.mcgstatus & MCG_STATUS_LMCES; |
1484 | |
1485 | /* |
1486 | * Local machine check may already know that we have to panic. |
1487 | * Broadcast machine check begins rendezvous in mce_start() |
1488 | * Go through all banks in exclusion of the other CPUs. This way we |
1489 | * don't report duplicated events on shared banks because the first one |
1490 | * to see it will clear it. |
1491 | */ |
1492 | if (lmce) { |
1493 | if (no_way_out) |
1494 | mce_panic(msg: "Fatal local machine check" , final: &m, exp: msg); |
1495 | } else { |
1496 | order = mce_start(no_way_out: &no_way_out); |
1497 | } |
1498 | |
1499 | taint = __mc_scan_banks(m: &m, regs, final, toclear, valid_banks, no_way_out, worst: &worst); |
1500 | |
1501 | if (!no_way_out) |
1502 | mce_clear_state(toclear); |
1503 | |
1504 | /* |
1505 | * Do most of the synchronization with other CPUs. |
1506 | * When there's any problem use only local no_way_out state. |
1507 | */ |
1508 | if (!lmce) { |
1509 | if (mce_end(order) < 0) { |
1510 | if (!no_way_out) |
1511 | no_way_out = worst >= MCE_PANIC_SEVERITY; |
1512 | |
1513 | if (no_way_out) |
1514 | mce_panic(msg: "Fatal machine check on current CPU" , final: &m, exp: msg); |
1515 | } |
1516 | } else { |
1517 | /* |
1518 | * If there was a fatal machine check we should have |
1519 | * already called mce_panic earlier in this function. |
1520 | * Since we re-read the banks, we might have found |
1521 | * something new. Check again to see if we found a |
1522 | * fatal error. We call "mce_severity()" again to |
1523 | * make sure we have the right "msg". |
1524 | */ |
1525 | if (worst >= MCE_PANIC_SEVERITY) { |
1526 | mce_severity(a: &m, regs, msg: &msg, is_excp: true); |
1527 | mce_panic(msg: "Local fatal machine check!" , final: &m, exp: msg); |
1528 | } |
1529 | } |
1530 | |
1531 | /* |
1532 | * Enable instrumentation around the external facilities like task_work_add() |
1533 | * (via queue_task_work()), fixup_exception() etc. For now, that is. Fixing this |
1534 | * properly would need a lot more involved reorganization. |
1535 | */ |
1536 | instrumentation_begin(); |
1537 | |
1538 | if (taint) |
1539 | add_taint(TAINT_MACHINE_CHECK, LOCKDEP_NOW_UNRELIABLE); |
1540 | |
1541 | if (worst != MCE_AR_SEVERITY && !kill_current_task) |
1542 | goto out; |
1543 | |
1544 | /* Fault was in user mode and we need to take some action */ |
1545 | if ((m.cs & 3) == 3) { |
1546 | /* If this triggers there is no way to recover. Die hard. */ |
1547 | BUG_ON(!on_thread_stack() || !user_mode(regs)); |
1548 | |
1549 | if (!mce_usable_address(&m)) |
1550 | queue_task_work(m: &m, msg, func: kill_me_now); |
1551 | else |
1552 | queue_task_work(m: &m, msg, func: kill_me_maybe); |
1553 | |
1554 | } else { |
1555 | /* |
1556 | * Handle an MCE which has happened in kernel space but from |
1557 | * which the kernel can recover: ex_has_fault_handler() has |
1558 | * already verified that the rIP at which the error happened is |
1559 | * a rIP from which the kernel can recover (by jumping to |
1560 | * recovery code specified in _ASM_EXTABLE_FAULT()) and the |
1561 | * corresponding exception handler which would do that is the |
1562 | * proper one. |
1563 | */ |
1564 | if (m.kflags & MCE_IN_KERNEL_RECOV) { |
1565 | if (!fixup_exception(regs, X86_TRAP_MC, error_code: 0, fault_addr: 0)) |
1566 | mce_panic(msg: "Failed kernel mode recovery" , final: &m, exp: msg); |
1567 | } |
1568 | |
1569 | if (m.kflags & MCE_IN_KERNEL_COPYIN) |
1570 | queue_task_work(m: &m, msg, func: kill_me_never); |
1571 | } |
1572 | |
1573 | out: |
1574 | instrumentation_end(); |
1575 | |
1576 | clear: |
1577 | mce_wrmsrl(MSR_IA32_MCG_STATUS, v: 0); |
1578 | } |
1579 | EXPORT_SYMBOL_GPL(do_machine_check); |
1580 | |
1581 | #ifndef CONFIG_MEMORY_FAILURE |
1582 | int memory_failure(unsigned long pfn, int flags) |
1583 | { |
1584 | /* mce_severity() should not hand us an ACTION_REQUIRED error */ |
1585 | BUG_ON(flags & MF_ACTION_REQUIRED); |
1586 | pr_err("Uncorrected memory error in page 0x%lx ignored\n" |
1587 | "Rebuild kernel with CONFIG_MEMORY_FAILURE=y for smarter handling\n" , |
1588 | pfn); |
1589 | |
1590 | return 0; |
1591 | } |
1592 | #endif |
1593 | |
1594 | /* |
1595 | * Periodic polling timer for "silent" machine check errors. If the |
1596 | * poller finds an MCE, poll 2x faster. When the poller finds no more |
1597 | * errors, poll 2x slower (up to check_interval seconds). |
1598 | */ |
1599 | static unsigned long check_interval = INITIAL_CHECK_INTERVAL; |
1600 | |
1601 | static DEFINE_PER_CPU(unsigned long, mce_next_interval); /* in jiffies */ |
1602 | static DEFINE_PER_CPU(struct timer_list, mce_timer); |
1603 | |
1604 | static unsigned long mce_adjust_timer_default(unsigned long interval) |
1605 | { |
1606 | return interval; |
1607 | } |
1608 | |
1609 | static unsigned long (*mce_adjust_timer)(unsigned long interval) = mce_adjust_timer_default; |
1610 | |
1611 | static void __start_timer(struct timer_list *t, unsigned long interval) |
1612 | { |
1613 | unsigned long when = jiffies + interval; |
1614 | unsigned long flags; |
1615 | |
1616 | local_irq_save(flags); |
1617 | |
1618 | if (!timer_pending(timer: t) || time_before(when, t->expires)) |
1619 | mod_timer(timer: t, expires: round_jiffies(j: when)); |
1620 | |
1621 | local_irq_restore(flags); |
1622 | } |
1623 | |
1624 | static void mc_poll_banks_default(void) |
1625 | { |
1626 | machine_check_poll(0, this_cpu_ptr(&mce_poll_banks)); |
1627 | } |
1628 | |
1629 | void (*mc_poll_banks)(void) = mc_poll_banks_default; |
1630 | |
1631 | static void mce_timer_fn(struct timer_list *t) |
1632 | { |
1633 | struct timer_list *cpu_t = this_cpu_ptr(&mce_timer); |
1634 | unsigned long iv; |
1635 | |
1636 | WARN_ON(cpu_t != t); |
1637 | |
1638 | iv = __this_cpu_read(mce_next_interval); |
1639 | |
1640 | if (mce_available(this_cpu_ptr(&cpu_info))) { |
1641 | mc_poll_banks(); |
1642 | |
1643 | if (mce_intel_cmci_poll()) { |
1644 | iv = mce_adjust_timer(iv); |
1645 | goto done; |
1646 | } |
1647 | } |
1648 | |
1649 | /* |
1650 | * Alert userspace if needed. If we logged an MCE, reduce the polling |
1651 | * interval, otherwise increase the polling interval. |
1652 | */ |
1653 | if (mce_notify_irq()) |
1654 | iv = max(iv / 2, (unsigned long) HZ/100); |
1655 | else |
1656 | iv = min(iv * 2, round_jiffies_relative(check_interval * HZ)); |
1657 | |
1658 | done: |
1659 | __this_cpu_write(mce_next_interval, iv); |
1660 | __start_timer(t, interval: iv); |
1661 | } |
1662 | |
1663 | /* |
1664 | * Ensure that the timer is firing in @interval from now. |
1665 | */ |
1666 | void mce_timer_kick(unsigned long interval) |
1667 | { |
1668 | struct timer_list *t = this_cpu_ptr(&mce_timer); |
1669 | unsigned long iv = __this_cpu_read(mce_next_interval); |
1670 | |
1671 | __start_timer(t, interval); |
1672 | |
1673 | if (interval < iv) |
1674 | __this_cpu_write(mce_next_interval, interval); |
1675 | } |
1676 | |
1677 | /* Must not be called in IRQ context where del_timer_sync() can deadlock */ |
1678 | static void mce_timer_delete_all(void) |
1679 | { |
1680 | int cpu; |
1681 | |
1682 | for_each_online_cpu(cpu) |
1683 | del_timer_sync(timer: &per_cpu(mce_timer, cpu)); |
1684 | } |
1685 | |
1686 | /* |
1687 | * Notify the user(s) about new machine check events. |
1688 | * Can be called from interrupt context, but not from machine check/NMI |
1689 | * context. |
1690 | */ |
1691 | int mce_notify_irq(void) |
1692 | { |
1693 | /* Not more than two messages every minute */ |
1694 | static DEFINE_RATELIMIT_STATE(ratelimit, 60*HZ, 2); |
1695 | |
1696 | if (test_and_clear_bit(nr: 0, addr: &mce_need_notify)) { |
1697 | mce_work_trigger(); |
1698 | |
1699 | if (__ratelimit(&ratelimit)) |
1700 | pr_info(HW_ERR "Machine check events logged\n" ); |
1701 | |
1702 | return 1; |
1703 | } |
1704 | return 0; |
1705 | } |
1706 | EXPORT_SYMBOL_GPL(mce_notify_irq); |
1707 | |
1708 | static void __mcheck_cpu_mce_banks_init(void) |
1709 | { |
1710 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
1711 | u8 n_banks = this_cpu_read(mce_num_banks); |
1712 | int i; |
1713 | |
1714 | for (i = 0; i < n_banks; i++) { |
1715 | struct mce_bank *b = &mce_banks[i]; |
1716 | |
1717 | /* |
1718 | * Init them all, __mcheck_cpu_apply_quirks() is going to apply |
1719 | * the required vendor quirks before |
1720 | * __mcheck_cpu_init_clear_banks() does the final bank setup. |
1721 | */ |
1722 | b->ctl = -1ULL; |
1723 | b->init = true; |
1724 | } |
1725 | } |
1726 | |
1727 | /* |
1728 | * Initialize Machine Checks for a CPU. |
1729 | */ |
1730 | static void __mcheck_cpu_cap_init(void) |
1731 | { |
1732 | u64 cap; |
1733 | u8 b; |
1734 | |
1735 | rdmsrl(MSR_IA32_MCG_CAP, cap); |
1736 | |
1737 | b = cap & MCG_BANKCNT_MASK; |
1738 | |
1739 | if (b > MAX_NR_BANKS) { |
1740 | pr_warn("CPU%d: Using only %u machine check banks out of %u\n" , |
1741 | smp_processor_id(), MAX_NR_BANKS, b); |
1742 | b = MAX_NR_BANKS; |
1743 | } |
1744 | |
1745 | this_cpu_write(mce_num_banks, b); |
1746 | |
1747 | __mcheck_cpu_mce_banks_init(); |
1748 | |
1749 | /* Use accurate RIP reporting if available. */ |
1750 | if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9) |
1751 | mca_cfg.rip_msr = MSR_IA32_MCG_EIP; |
1752 | |
1753 | if (cap & MCG_SER_P) |
1754 | mca_cfg.ser = 1; |
1755 | } |
1756 | |
1757 | static void __mcheck_cpu_init_generic(void) |
1758 | { |
1759 | enum mcp_flags m_fl = 0; |
1760 | mce_banks_t all_banks; |
1761 | u64 cap; |
1762 | |
1763 | if (!mca_cfg.bootlog) |
1764 | m_fl = MCP_DONTLOG; |
1765 | |
1766 | /* |
1767 | * Log the machine checks left over from the previous reset. Log them |
1768 | * only, do not start processing them. That will happen in mcheck_late_init() |
1769 | * when all consumers have been registered on the notifier chain. |
1770 | */ |
1771 | bitmap_fill(dst: all_banks, MAX_NR_BANKS); |
1772 | machine_check_poll(MCP_UC | MCP_QUEUE_LOG | m_fl, &all_banks); |
1773 | |
1774 | cr4_set_bits(X86_CR4_MCE); |
1775 | |
1776 | rdmsrl(MSR_IA32_MCG_CAP, cap); |
1777 | if (cap & MCG_CTL_P) |
1778 | wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff); |
1779 | } |
1780 | |
1781 | static void __mcheck_cpu_init_clear_banks(void) |
1782 | { |
1783 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
1784 | int i; |
1785 | |
1786 | for (i = 0; i < this_cpu_read(mce_num_banks); i++) { |
1787 | struct mce_bank *b = &mce_banks[i]; |
1788 | |
1789 | if (!b->init) |
1790 | continue; |
1791 | wrmsrl(msr: mca_msr_reg(bank: i, reg: MCA_CTL), val: b->ctl); |
1792 | wrmsrl(msr: mca_msr_reg(bank: i, reg: MCA_STATUS), val: 0); |
1793 | } |
1794 | } |
1795 | |
1796 | /* |
1797 | * Do a final check to see if there are any unused/RAZ banks. |
1798 | * |
1799 | * This must be done after the banks have been initialized and any quirks have |
1800 | * been applied. |
1801 | * |
1802 | * Do not call this from any user-initiated flows, e.g. CPU hotplug or sysfs. |
1803 | * Otherwise, a user who disables a bank will not be able to re-enable it |
1804 | * without a system reboot. |
1805 | */ |
1806 | static void __mcheck_cpu_check_banks(void) |
1807 | { |
1808 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
1809 | u64 msrval; |
1810 | int i; |
1811 | |
1812 | for (i = 0; i < this_cpu_read(mce_num_banks); i++) { |
1813 | struct mce_bank *b = &mce_banks[i]; |
1814 | |
1815 | if (!b->init) |
1816 | continue; |
1817 | |
1818 | rdmsrl(mca_msr_reg(i, MCA_CTL), msrval); |
1819 | b->init = !!msrval; |
1820 | } |
1821 | } |
1822 | |
1823 | /* Add per CPU specific workarounds here */ |
1824 | static int __mcheck_cpu_apply_quirks(struct cpuinfo_x86 *c) |
1825 | { |
1826 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
1827 | struct mca_config *cfg = &mca_cfg; |
1828 | |
1829 | if (c->x86_vendor == X86_VENDOR_UNKNOWN) { |
1830 | pr_info("unknown CPU type - not enabling MCE support\n" ); |
1831 | return -EOPNOTSUPP; |
1832 | } |
1833 | |
1834 | /* This should be disabled by the BIOS, but isn't always */ |
1835 | if (c->x86_vendor == X86_VENDOR_AMD) { |
1836 | if (c->x86 == 15 && this_cpu_read(mce_num_banks) > 4) { |
1837 | /* |
1838 | * disable GART TBL walk error reporting, which |
1839 | * trips off incorrectly with the IOMMU & 3ware |
1840 | * & Cerberus: |
1841 | */ |
1842 | clear_bit(nr: 10, addr: (unsigned long *)&mce_banks[4].ctl); |
1843 | } |
1844 | if (c->x86 < 0x11 && cfg->bootlog < 0) { |
1845 | /* |
1846 | * Lots of broken BIOS around that don't clear them |
1847 | * by default and leave crap in there. Don't log: |
1848 | */ |
1849 | cfg->bootlog = 0; |
1850 | } |
1851 | /* |
1852 | * Various K7s with broken bank 0 around. Always disable |
1853 | * by default. |
1854 | */ |
1855 | if (c->x86 == 6 && this_cpu_read(mce_num_banks) > 0) |
1856 | mce_banks[0].ctl = 0; |
1857 | |
1858 | /* |
1859 | * overflow_recov is supported for F15h Models 00h-0fh |
1860 | * even though we don't have a CPUID bit for it. |
1861 | */ |
1862 | if (c->x86 == 0x15 && c->x86_model <= 0xf) |
1863 | mce_flags.overflow_recov = 1; |
1864 | |
1865 | if (c->x86 >= 0x17 && c->x86 <= 0x1A) |
1866 | mce_flags.zen_ifu_quirk = 1; |
1867 | |
1868 | } |
1869 | |
1870 | if (c->x86_vendor == X86_VENDOR_INTEL) { |
1871 | /* |
1872 | * SDM documents that on family 6 bank 0 should not be written |
1873 | * because it aliases to another special BIOS controlled |
1874 | * register. |
1875 | * But it's not aliased anymore on model 0x1a+ |
1876 | * Don't ignore bank 0 completely because there could be a |
1877 | * valid event later, merely don't write CTL0. |
1878 | */ |
1879 | |
1880 | if (c->x86 == 6 && c->x86_model < 0x1A && this_cpu_read(mce_num_banks) > 0) |
1881 | mce_banks[0].init = false; |
1882 | |
1883 | /* |
1884 | * All newer Intel systems support MCE broadcasting. Enable |
1885 | * synchronization with a one second timeout. |
1886 | */ |
1887 | if ((c->x86 > 6 || (c->x86 == 6 && c->x86_model >= 0xe)) && |
1888 | cfg->monarch_timeout < 0) |
1889 | cfg->monarch_timeout = USEC_PER_SEC; |
1890 | |
1891 | /* |
1892 | * There are also broken BIOSes on some Pentium M and |
1893 | * earlier systems: |
1894 | */ |
1895 | if (c->x86 == 6 && c->x86_model <= 13 && cfg->bootlog < 0) |
1896 | cfg->bootlog = 0; |
1897 | |
1898 | if (c->x86 == 6 && c->x86_model == 45) |
1899 | mce_flags.snb_ifu_quirk = 1; |
1900 | |
1901 | /* |
1902 | * Skylake, Cascacde Lake and Cooper Lake require a quirk on |
1903 | * rep movs. |
1904 | */ |
1905 | if (c->x86 == 6 && c->x86_model == INTEL_FAM6_SKYLAKE_X) |
1906 | mce_flags.skx_repmov_quirk = 1; |
1907 | } |
1908 | |
1909 | if (c->x86_vendor == X86_VENDOR_ZHAOXIN) { |
1910 | /* |
1911 | * All newer Zhaoxin CPUs support MCE broadcasting. Enable |
1912 | * synchronization with a one second timeout. |
1913 | */ |
1914 | if (c->x86 > 6 || (c->x86_model == 0x19 || c->x86_model == 0x1f)) { |
1915 | if (cfg->monarch_timeout < 0) |
1916 | cfg->monarch_timeout = USEC_PER_SEC; |
1917 | } |
1918 | } |
1919 | |
1920 | if (cfg->monarch_timeout < 0) |
1921 | cfg->monarch_timeout = 0; |
1922 | if (cfg->bootlog != 0) |
1923 | cfg->panic_timeout = 30; |
1924 | |
1925 | return 0; |
1926 | } |
1927 | |
1928 | static int __mcheck_cpu_ancient_init(struct cpuinfo_x86 *c) |
1929 | { |
1930 | if (c->x86 != 5) |
1931 | return 0; |
1932 | |
1933 | switch (c->x86_vendor) { |
1934 | case X86_VENDOR_INTEL: |
1935 | intel_p5_mcheck_init(c); |
1936 | mce_flags.p5 = 1; |
1937 | return 1; |
1938 | case X86_VENDOR_CENTAUR: |
1939 | winchip_mcheck_init(c); |
1940 | mce_flags.winchip = 1; |
1941 | return 1; |
1942 | default: |
1943 | return 0; |
1944 | } |
1945 | |
1946 | return 0; |
1947 | } |
1948 | |
1949 | /* |
1950 | * Init basic CPU features needed for early decoding of MCEs. |
1951 | */ |
1952 | static void __mcheck_cpu_init_early(struct cpuinfo_x86 *c) |
1953 | { |
1954 | if (c->x86_vendor == X86_VENDOR_AMD || c->x86_vendor == X86_VENDOR_HYGON) { |
1955 | mce_flags.overflow_recov = !!cpu_has(c, X86_FEATURE_OVERFLOW_RECOV); |
1956 | mce_flags.succor = !!cpu_has(c, X86_FEATURE_SUCCOR); |
1957 | mce_flags.smca = !!cpu_has(c, X86_FEATURE_SMCA); |
1958 | mce_flags.amd_threshold = 1; |
1959 | } |
1960 | } |
1961 | |
1962 | static void mce_centaur_feature_init(struct cpuinfo_x86 *c) |
1963 | { |
1964 | struct mca_config *cfg = &mca_cfg; |
1965 | |
1966 | /* |
1967 | * All newer Centaur CPUs support MCE broadcasting. Enable |
1968 | * synchronization with a one second timeout. |
1969 | */ |
1970 | if ((c->x86 == 6 && c->x86_model == 0xf && c->x86_stepping >= 0xe) || |
1971 | c->x86 > 6) { |
1972 | if (cfg->monarch_timeout < 0) |
1973 | cfg->monarch_timeout = USEC_PER_SEC; |
1974 | } |
1975 | } |
1976 | |
1977 | static void mce_zhaoxin_feature_init(struct cpuinfo_x86 *c) |
1978 | { |
1979 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
1980 | |
1981 | /* |
1982 | * These CPUs have MCA bank 8 which reports only one error type called |
1983 | * SVAD (System View Address Decoder). The reporting of that error is |
1984 | * controlled by IA32_MC8.CTL.0. |
1985 | * |
1986 | * If enabled, prefetching on these CPUs will cause SVAD MCE when |
1987 | * virtual machines start and result in a system panic. Always disable |
1988 | * bank 8 SVAD error by default. |
1989 | */ |
1990 | if ((c->x86 == 7 && c->x86_model == 0x1b) || |
1991 | (c->x86_model == 0x19 || c->x86_model == 0x1f)) { |
1992 | if (this_cpu_read(mce_num_banks) > 8) |
1993 | mce_banks[8].ctl = 0; |
1994 | } |
1995 | |
1996 | intel_init_cmci(); |
1997 | intel_init_lmce(); |
1998 | mce_adjust_timer = cmci_intel_adjust_timer; |
1999 | } |
2000 | |
2001 | static void mce_zhaoxin_feature_clear(struct cpuinfo_x86 *c) |
2002 | { |
2003 | intel_clear_lmce(); |
2004 | } |
2005 | |
2006 | static void __mcheck_cpu_init_vendor(struct cpuinfo_x86 *c) |
2007 | { |
2008 | switch (c->x86_vendor) { |
2009 | case X86_VENDOR_INTEL: |
2010 | mce_intel_feature_init(c); |
2011 | mce_adjust_timer = cmci_intel_adjust_timer; |
2012 | break; |
2013 | |
2014 | case X86_VENDOR_AMD: { |
2015 | mce_amd_feature_init(c); |
2016 | break; |
2017 | } |
2018 | |
2019 | case X86_VENDOR_HYGON: |
2020 | mce_hygon_feature_init(c); |
2021 | break; |
2022 | |
2023 | case X86_VENDOR_CENTAUR: |
2024 | mce_centaur_feature_init(c); |
2025 | break; |
2026 | |
2027 | case X86_VENDOR_ZHAOXIN: |
2028 | mce_zhaoxin_feature_init(c); |
2029 | break; |
2030 | |
2031 | default: |
2032 | break; |
2033 | } |
2034 | } |
2035 | |
2036 | static void __mcheck_cpu_clear_vendor(struct cpuinfo_x86 *c) |
2037 | { |
2038 | switch (c->x86_vendor) { |
2039 | case X86_VENDOR_INTEL: |
2040 | mce_intel_feature_clear(c); |
2041 | break; |
2042 | |
2043 | case X86_VENDOR_ZHAOXIN: |
2044 | mce_zhaoxin_feature_clear(c); |
2045 | break; |
2046 | |
2047 | default: |
2048 | break; |
2049 | } |
2050 | } |
2051 | |
2052 | static void mce_start_timer(struct timer_list *t) |
2053 | { |
2054 | unsigned long iv = check_interval * HZ; |
2055 | |
2056 | if (mca_cfg.ignore_ce || !iv) |
2057 | return; |
2058 | |
2059 | this_cpu_write(mce_next_interval, iv); |
2060 | __start_timer(t, interval: iv); |
2061 | } |
2062 | |
2063 | static void __mcheck_cpu_setup_timer(void) |
2064 | { |
2065 | struct timer_list *t = this_cpu_ptr(&mce_timer); |
2066 | |
2067 | timer_setup(t, mce_timer_fn, TIMER_PINNED); |
2068 | } |
2069 | |
2070 | static void __mcheck_cpu_init_timer(void) |
2071 | { |
2072 | struct timer_list *t = this_cpu_ptr(&mce_timer); |
2073 | |
2074 | timer_setup(t, mce_timer_fn, TIMER_PINNED); |
2075 | mce_start_timer(t); |
2076 | } |
2077 | |
2078 | bool filter_mce(struct mce *m) |
2079 | { |
2080 | if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) |
2081 | return amd_filter_mce(m); |
2082 | if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) |
2083 | return intel_filter_mce(m); |
2084 | |
2085 | return false; |
2086 | } |
2087 | |
2088 | static __always_inline void exc_machine_check_kernel(struct pt_regs *regs) |
2089 | { |
2090 | irqentry_state_t irq_state; |
2091 | |
2092 | WARN_ON_ONCE(user_mode(regs)); |
2093 | |
2094 | /* |
2095 | * Only required when from kernel mode. See |
2096 | * mce_check_crashing_cpu() for details. |
2097 | */ |
2098 | if (mca_cfg.initialized && mce_check_crashing_cpu()) |
2099 | return; |
2100 | |
2101 | irq_state = irqentry_nmi_enter(regs); |
2102 | |
2103 | do_machine_check(regs); |
2104 | |
2105 | irqentry_nmi_exit(regs, irq_state); |
2106 | } |
2107 | |
2108 | static __always_inline void exc_machine_check_user(struct pt_regs *regs) |
2109 | { |
2110 | irqentry_enter_from_user_mode(regs); |
2111 | |
2112 | do_machine_check(regs); |
2113 | |
2114 | irqentry_exit_to_user_mode(regs); |
2115 | } |
2116 | |
2117 | #ifdef CONFIG_X86_64 |
2118 | /* MCE hit kernel mode */ |
2119 | DEFINE_IDTENTRY_MCE(exc_machine_check) |
2120 | { |
2121 | unsigned long dr7; |
2122 | |
2123 | dr7 = local_db_save(); |
2124 | exc_machine_check_kernel(regs); |
2125 | local_db_restore(dr7); |
2126 | } |
2127 | |
2128 | /* The user mode variant. */ |
2129 | DEFINE_IDTENTRY_MCE_USER(exc_machine_check) |
2130 | { |
2131 | unsigned long dr7; |
2132 | |
2133 | dr7 = local_db_save(); |
2134 | exc_machine_check_user(regs); |
2135 | local_db_restore(dr7); |
2136 | } |
2137 | #else |
2138 | /* 32bit unified entry point */ |
2139 | DEFINE_IDTENTRY_RAW(exc_machine_check) |
2140 | { |
2141 | unsigned long dr7; |
2142 | |
2143 | dr7 = local_db_save(); |
2144 | if (user_mode(regs)) |
2145 | exc_machine_check_user(regs); |
2146 | else |
2147 | exc_machine_check_kernel(regs); |
2148 | local_db_restore(dr7); |
2149 | } |
2150 | #endif |
2151 | |
2152 | /* |
2153 | * Called for each booted CPU to set up machine checks. |
2154 | * Must be called with preempt off: |
2155 | */ |
2156 | void mcheck_cpu_init(struct cpuinfo_x86 *c) |
2157 | { |
2158 | if (mca_cfg.disabled) |
2159 | return; |
2160 | |
2161 | if (__mcheck_cpu_ancient_init(c)) |
2162 | return; |
2163 | |
2164 | if (!mce_available(c)) |
2165 | return; |
2166 | |
2167 | __mcheck_cpu_cap_init(); |
2168 | |
2169 | if (__mcheck_cpu_apply_quirks(c) < 0) { |
2170 | mca_cfg.disabled = 1; |
2171 | return; |
2172 | } |
2173 | |
2174 | if (mce_gen_pool_init()) { |
2175 | mca_cfg.disabled = 1; |
2176 | pr_emerg("Couldn't allocate MCE records pool!\n" ); |
2177 | return; |
2178 | } |
2179 | |
2180 | mca_cfg.initialized = 1; |
2181 | |
2182 | __mcheck_cpu_init_early(c); |
2183 | __mcheck_cpu_init_generic(); |
2184 | __mcheck_cpu_init_vendor(c); |
2185 | __mcheck_cpu_init_clear_banks(); |
2186 | __mcheck_cpu_check_banks(); |
2187 | __mcheck_cpu_setup_timer(); |
2188 | } |
2189 | |
2190 | /* |
2191 | * Called for each booted CPU to clear some machine checks opt-ins |
2192 | */ |
2193 | void mcheck_cpu_clear(struct cpuinfo_x86 *c) |
2194 | { |
2195 | if (mca_cfg.disabled) |
2196 | return; |
2197 | |
2198 | if (!mce_available(c)) |
2199 | return; |
2200 | |
2201 | /* |
2202 | * Possibly to clear general settings generic to x86 |
2203 | * __mcheck_cpu_clear_generic(c); |
2204 | */ |
2205 | __mcheck_cpu_clear_vendor(c); |
2206 | |
2207 | } |
2208 | |
2209 | static void __mce_disable_bank(void *arg) |
2210 | { |
2211 | int bank = *((int *)arg); |
2212 | __clear_bit(bank, this_cpu_ptr(mce_poll_banks)); |
2213 | cmci_disable_bank(bank); |
2214 | } |
2215 | |
2216 | void mce_disable_bank(int bank) |
2217 | { |
2218 | if (bank >= this_cpu_read(mce_num_banks)) { |
2219 | pr_warn(FW_BUG |
2220 | "Ignoring request to disable invalid MCA bank %d.\n" , |
2221 | bank); |
2222 | return; |
2223 | } |
2224 | set_bit(nr: bank, addr: mce_banks_ce_disabled); |
2225 | on_each_cpu(func: __mce_disable_bank, info: &bank, wait: 1); |
2226 | } |
2227 | |
2228 | /* |
2229 | * mce=off Disables machine check |
2230 | * mce=no_cmci Disables CMCI |
2231 | * mce=no_lmce Disables LMCE |
2232 | * mce=dont_log_ce Clears corrected events silently, no log created for CEs. |
2233 | * mce=print_all Print all machine check logs to console |
2234 | * mce=ignore_ce Disables polling and CMCI, corrected events are not cleared. |
2235 | * mce=TOLERANCELEVEL[,monarchtimeout] (number, see above) |
2236 | * monarchtimeout is how long to wait for other CPUs on machine |
2237 | * check, or 0 to not wait |
2238 | * mce=bootlog Log MCEs from before booting. Disabled by default on AMD Fam10h |
2239 | and older. |
2240 | * mce=nobootlog Don't log MCEs from before booting. |
2241 | * mce=bios_cmci_threshold Don't program the CMCI threshold |
2242 | * mce=recovery force enable copy_mc_fragile() |
2243 | */ |
2244 | static int __init mcheck_enable(char *str) |
2245 | { |
2246 | struct mca_config *cfg = &mca_cfg; |
2247 | |
2248 | if (*str == 0) { |
2249 | enable_p5_mce(); |
2250 | return 1; |
2251 | } |
2252 | if (*str == '=') |
2253 | str++; |
2254 | if (!strcmp(str, "off" )) |
2255 | cfg->disabled = 1; |
2256 | else if (!strcmp(str, "no_cmci" )) |
2257 | cfg->cmci_disabled = true; |
2258 | else if (!strcmp(str, "no_lmce" )) |
2259 | cfg->lmce_disabled = 1; |
2260 | else if (!strcmp(str, "dont_log_ce" )) |
2261 | cfg->dont_log_ce = true; |
2262 | else if (!strcmp(str, "print_all" )) |
2263 | cfg->print_all = true; |
2264 | else if (!strcmp(str, "ignore_ce" )) |
2265 | cfg->ignore_ce = true; |
2266 | else if (!strcmp(str, "bootlog" ) || !strcmp(str, "nobootlog" )) |
2267 | cfg->bootlog = (str[0] == 'b'); |
2268 | else if (!strcmp(str, "bios_cmci_threshold" )) |
2269 | cfg->bios_cmci_threshold = 1; |
2270 | else if (!strcmp(str, "recovery" )) |
2271 | cfg->recovery = 1; |
2272 | else if (isdigit(c: str[0])) |
2273 | get_option(str: &str, pint: &(cfg->monarch_timeout)); |
2274 | else { |
2275 | pr_info("mce argument %s ignored. Please use /sys\n" , str); |
2276 | return 0; |
2277 | } |
2278 | return 1; |
2279 | } |
2280 | __setup("mce" , mcheck_enable); |
2281 | |
2282 | int __init mcheck_init(void) |
2283 | { |
2284 | mce_register_decode_chain(&early_nb); |
2285 | mce_register_decode_chain(&mce_uc_nb); |
2286 | mce_register_decode_chain(&mce_default_nb); |
2287 | |
2288 | INIT_WORK(&mce_work, mce_gen_pool_process); |
2289 | init_irq_work(work: &mce_irq_work, func: mce_irq_work_cb); |
2290 | |
2291 | return 0; |
2292 | } |
2293 | |
2294 | /* |
2295 | * mce_syscore: PM support |
2296 | */ |
2297 | |
2298 | /* |
2299 | * Disable machine checks on suspend and shutdown. We can't really handle |
2300 | * them later. |
2301 | */ |
2302 | static void mce_disable_error_reporting(void) |
2303 | { |
2304 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
2305 | int i; |
2306 | |
2307 | for (i = 0; i < this_cpu_read(mce_num_banks); i++) { |
2308 | struct mce_bank *b = &mce_banks[i]; |
2309 | |
2310 | if (b->init) |
2311 | wrmsrl(msr: mca_msr_reg(bank: i, reg: MCA_CTL), val: 0); |
2312 | } |
2313 | return; |
2314 | } |
2315 | |
2316 | static void vendor_disable_error_reporting(void) |
2317 | { |
2318 | /* |
2319 | * Don't clear on Intel or AMD or Hygon or Zhaoxin CPUs. Some of these |
2320 | * MSRs are socket-wide. Disabling them for just a single offlined CPU |
2321 | * is bad, since it will inhibit reporting for all shared resources on |
2322 | * the socket like the last level cache (LLC), the integrated memory |
2323 | * controller (iMC), etc. |
2324 | */ |
2325 | if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL || |
2326 | boot_cpu_data.x86_vendor == X86_VENDOR_HYGON || |
2327 | boot_cpu_data.x86_vendor == X86_VENDOR_AMD || |
2328 | boot_cpu_data.x86_vendor == X86_VENDOR_ZHAOXIN) |
2329 | return; |
2330 | |
2331 | mce_disable_error_reporting(); |
2332 | } |
2333 | |
2334 | static int mce_syscore_suspend(void) |
2335 | { |
2336 | vendor_disable_error_reporting(); |
2337 | return 0; |
2338 | } |
2339 | |
2340 | static void mce_syscore_shutdown(void) |
2341 | { |
2342 | vendor_disable_error_reporting(); |
2343 | } |
2344 | |
2345 | /* |
2346 | * On resume clear all MCE state. Don't want to see leftovers from the BIOS. |
2347 | * Only one CPU is active at this time, the others get re-added later using |
2348 | * CPU hotplug: |
2349 | */ |
2350 | static void mce_syscore_resume(void) |
2351 | { |
2352 | __mcheck_cpu_init_generic(); |
2353 | __mcheck_cpu_init_vendor(raw_cpu_ptr(&cpu_info)); |
2354 | __mcheck_cpu_init_clear_banks(); |
2355 | } |
2356 | |
2357 | static struct syscore_ops mce_syscore_ops = { |
2358 | .suspend = mce_syscore_suspend, |
2359 | .shutdown = mce_syscore_shutdown, |
2360 | .resume = mce_syscore_resume, |
2361 | }; |
2362 | |
2363 | /* |
2364 | * mce_device: Sysfs support |
2365 | */ |
2366 | |
2367 | static void mce_cpu_restart(void *data) |
2368 | { |
2369 | if (!mce_available(raw_cpu_ptr(&cpu_info))) |
2370 | return; |
2371 | __mcheck_cpu_init_generic(); |
2372 | __mcheck_cpu_init_clear_banks(); |
2373 | __mcheck_cpu_init_timer(); |
2374 | } |
2375 | |
2376 | /* Reinit MCEs after user configuration changes */ |
2377 | static void mce_restart(void) |
2378 | { |
2379 | mce_timer_delete_all(); |
2380 | on_each_cpu(func: mce_cpu_restart, NULL, wait: 1); |
2381 | mce_schedule_work(); |
2382 | } |
2383 | |
2384 | /* Toggle features for corrected errors */ |
2385 | static void mce_disable_cmci(void *data) |
2386 | { |
2387 | if (!mce_available(raw_cpu_ptr(&cpu_info))) |
2388 | return; |
2389 | cmci_clear(); |
2390 | } |
2391 | |
2392 | static void mce_enable_ce(void *all) |
2393 | { |
2394 | if (!mce_available(raw_cpu_ptr(&cpu_info))) |
2395 | return; |
2396 | cmci_reenable(); |
2397 | cmci_recheck(); |
2398 | if (all) |
2399 | __mcheck_cpu_init_timer(); |
2400 | } |
2401 | |
2402 | static struct bus_type mce_subsys = { |
2403 | .name = "machinecheck" , |
2404 | .dev_name = "machinecheck" , |
2405 | }; |
2406 | |
2407 | DEFINE_PER_CPU(struct device *, mce_device); |
2408 | |
2409 | static inline struct mce_bank_dev *attr_to_bank(struct device_attribute *attr) |
2410 | { |
2411 | return container_of(attr, struct mce_bank_dev, attr); |
2412 | } |
2413 | |
2414 | static ssize_t show_bank(struct device *s, struct device_attribute *attr, |
2415 | char *buf) |
2416 | { |
2417 | u8 bank = attr_to_bank(attr)->bank; |
2418 | struct mce_bank *b; |
2419 | |
2420 | if (bank >= per_cpu(mce_num_banks, s->id)) |
2421 | return -EINVAL; |
2422 | |
2423 | b = &per_cpu(mce_banks_array, s->id)[bank]; |
2424 | |
2425 | if (!b->init) |
2426 | return -ENODEV; |
2427 | |
2428 | return sprintf(buf, fmt: "%llx\n" , b->ctl); |
2429 | } |
2430 | |
2431 | static ssize_t set_bank(struct device *s, struct device_attribute *attr, |
2432 | const char *buf, size_t size) |
2433 | { |
2434 | u8 bank = attr_to_bank(attr)->bank; |
2435 | struct mce_bank *b; |
2436 | u64 new; |
2437 | |
2438 | if (kstrtou64(s: buf, base: 0, res: &new) < 0) |
2439 | return -EINVAL; |
2440 | |
2441 | if (bank >= per_cpu(mce_num_banks, s->id)) |
2442 | return -EINVAL; |
2443 | |
2444 | b = &per_cpu(mce_banks_array, s->id)[bank]; |
2445 | |
2446 | if (!b->init) |
2447 | return -ENODEV; |
2448 | |
2449 | b->ctl = new; |
2450 | mce_restart(); |
2451 | |
2452 | return size; |
2453 | } |
2454 | |
2455 | static ssize_t set_ignore_ce(struct device *s, |
2456 | struct device_attribute *attr, |
2457 | const char *buf, size_t size) |
2458 | { |
2459 | u64 new; |
2460 | |
2461 | if (kstrtou64(s: buf, base: 0, res: &new) < 0) |
2462 | return -EINVAL; |
2463 | |
2464 | mutex_lock(&mce_sysfs_mutex); |
2465 | if (mca_cfg.ignore_ce ^ !!new) { |
2466 | if (new) { |
2467 | /* disable ce features */ |
2468 | mce_timer_delete_all(); |
2469 | on_each_cpu(func: mce_disable_cmci, NULL, wait: 1); |
2470 | mca_cfg.ignore_ce = true; |
2471 | } else { |
2472 | /* enable ce features */ |
2473 | mca_cfg.ignore_ce = false; |
2474 | on_each_cpu(func: mce_enable_ce, info: (void *)1, wait: 1); |
2475 | } |
2476 | } |
2477 | mutex_unlock(lock: &mce_sysfs_mutex); |
2478 | |
2479 | return size; |
2480 | } |
2481 | |
2482 | static ssize_t set_cmci_disabled(struct device *s, |
2483 | struct device_attribute *attr, |
2484 | const char *buf, size_t size) |
2485 | { |
2486 | u64 new; |
2487 | |
2488 | if (kstrtou64(s: buf, base: 0, res: &new) < 0) |
2489 | return -EINVAL; |
2490 | |
2491 | mutex_lock(&mce_sysfs_mutex); |
2492 | if (mca_cfg.cmci_disabled ^ !!new) { |
2493 | if (new) { |
2494 | /* disable cmci */ |
2495 | on_each_cpu(func: mce_disable_cmci, NULL, wait: 1); |
2496 | mca_cfg.cmci_disabled = true; |
2497 | } else { |
2498 | /* enable cmci */ |
2499 | mca_cfg.cmci_disabled = false; |
2500 | on_each_cpu(func: mce_enable_ce, NULL, wait: 1); |
2501 | } |
2502 | } |
2503 | mutex_unlock(lock: &mce_sysfs_mutex); |
2504 | |
2505 | return size; |
2506 | } |
2507 | |
2508 | static ssize_t store_int_with_restart(struct device *s, |
2509 | struct device_attribute *attr, |
2510 | const char *buf, size_t size) |
2511 | { |
2512 | unsigned long old_check_interval = check_interval; |
2513 | ssize_t ret = device_store_ulong(dev: s, attr, buf, count: size); |
2514 | |
2515 | if (check_interval == old_check_interval) |
2516 | return ret; |
2517 | |
2518 | mutex_lock(&mce_sysfs_mutex); |
2519 | mce_restart(); |
2520 | mutex_unlock(lock: &mce_sysfs_mutex); |
2521 | |
2522 | return ret; |
2523 | } |
2524 | |
2525 | static DEVICE_INT_ATTR(monarch_timeout, 0644, mca_cfg.monarch_timeout); |
2526 | static DEVICE_BOOL_ATTR(dont_log_ce, 0644, mca_cfg.dont_log_ce); |
2527 | static DEVICE_BOOL_ATTR(print_all, 0644, mca_cfg.print_all); |
2528 | |
2529 | static struct dev_ext_attribute dev_attr_check_interval = { |
2530 | __ATTR(check_interval, 0644, device_show_int, store_int_with_restart), |
2531 | &check_interval |
2532 | }; |
2533 | |
2534 | static struct dev_ext_attribute dev_attr_ignore_ce = { |
2535 | __ATTR(ignore_ce, 0644, device_show_bool, set_ignore_ce), |
2536 | &mca_cfg.ignore_ce |
2537 | }; |
2538 | |
2539 | static struct dev_ext_attribute dev_attr_cmci_disabled = { |
2540 | __ATTR(cmci_disabled, 0644, device_show_bool, set_cmci_disabled), |
2541 | &mca_cfg.cmci_disabled |
2542 | }; |
2543 | |
2544 | static struct device_attribute *mce_device_attrs[] = { |
2545 | &dev_attr_check_interval.attr, |
2546 | #ifdef CONFIG_X86_MCELOG_LEGACY |
2547 | &dev_attr_trigger, |
2548 | #endif |
2549 | &dev_attr_monarch_timeout.attr, |
2550 | &dev_attr_dont_log_ce.attr, |
2551 | &dev_attr_print_all.attr, |
2552 | &dev_attr_ignore_ce.attr, |
2553 | &dev_attr_cmci_disabled.attr, |
2554 | NULL |
2555 | }; |
2556 | |
2557 | static cpumask_var_t mce_device_initialized; |
2558 | |
2559 | static void mce_device_release(struct device *dev) |
2560 | { |
2561 | kfree(objp: dev); |
2562 | } |
2563 | |
2564 | /* Per CPU device init. All of the CPUs still share the same bank device: */ |
2565 | static int mce_device_create(unsigned int cpu) |
2566 | { |
2567 | struct device *dev; |
2568 | int err; |
2569 | int i, j; |
2570 | |
2571 | if (!mce_available(c: &boot_cpu_data)) |
2572 | return -EIO; |
2573 | |
2574 | dev = per_cpu(mce_device, cpu); |
2575 | if (dev) |
2576 | return 0; |
2577 | |
2578 | dev = kzalloc(size: sizeof(*dev), GFP_KERNEL); |
2579 | if (!dev) |
2580 | return -ENOMEM; |
2581 | dev->id = cpu; |
2582 | dev->bus = &mce_subsys; |
2583 | dev->release = &mce_device_release; |
2584 | |
2585 | err = device_register(dev); |
2586 | if (err) { |
2587 | put_device(dev); |
2588 | return err; |
2589 | } |
2590 | |
2591 | for (i = 0; mce_device_attrs[i]; i++) { |
2592 | err = device_create_file(device: dev, entry: mce_device_attrs[i]); |
2593 | if (err) |
2594 | goto error; |
2595 | } |
2596 | for (j = 0; j < per_cpu(mce_num_banks, cpu); j++) { |
2597 | err = device_create_file(device: dev, entry: &mce_bank_devs[j].attr); |
2598 | if (err) |
2599 | goto error2; |
2600 | } |
2601 | cpumask_set_cpu(cpu, dstp: mce_device_initialized); |
2602 | per_cpu(mce_device, cpu) = dev; |
2603 | |
2604 | return 0; |
2605 | error2: |
2606 | while (--j >= 0) |
2607 | device_remove_file(dev, attr: &mce_bank_devs[j].attr); |
2608 | error: |
2609 | while (--i >= 0) |
2610 | device_remove_file(dev, attr: mce_device_attrs[i]); |
2611 | |
2612 | device_unregister(dev); |
2613 | |
2614 | return err; |
2615 | } |
2616 | |
2617 | static void mce_device_remove(unsigned int cpu) |
2618 | { |
2619 | struct device *dev = per_cpu(mce_device, cpu); |
2620 | int i; |
2621 | |
2622 | if (!cpumask_test_cpu(cpu, cpumask: mce_device_initialized)) |
2623 | return; |
2624 | |
2625 | for (i = 0; mce_device_attrs[i]; i++) |
2626 | device_remove_file(dev, attr: mce_device_attrs[i]); |
2627 | |
2628 | for (i = 0; i < per_cpu(mce_num_banks, cpu); i++) |
2629 | device_remove_file(dev, attr: &mce_bank_devs[i].attr); |
2630 | |
2631 | device_unregister(dev); |
2632 | cpumask_clear_cpu(cpu, dstp: mce_device_initialized); |
2633 | per_cpu(mce_device, cpu) = NULL; |
2634 | } |
2635 | |
2636 | /* Make sure there are no machine checks on offlined CPUs. */ |
2637 | static void mce_disable_cpu(void) |
2638 | { |
2639 | if (!mce_available(raw_cpu_ptr(&cpu_info))) |
2640 | return; |
2641 | |
2642 | if (!cpuhp_tasks_frozen) |
2643 | cmci_clear(); |
2644 | |
2645 | vendor_disable_error_reporting(); |
2646 | } |
2647 | |
2648 | static void mce_reenable_cpu(void) |
2649 | { |
2650 | struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array); |
2651 | int i; |
2652 | |
2653 | if (!mce_available(raw_cpu_ptr(&cpu_info))) |
2654 | return; |
2655 | |
2656 | if (!cpuhp_tasks_frozen) |
2657 | cmci_reenable(); |
2658 | for (i = 0; i < this_cpu_read(mce_num_banks); i++) { |
2659 | struct mce_bank *b = &mce_banks[i]; |
2660 | |
2661 | if (b->init) |
2662 | wrmsrl(msr: mca_msr_reg(bank: i, reg: MCA_CTL), val: b->ctl); |
2663 | } |
2664 | } |
2665 | |
2666 | static int mce_cpu_dead(unsigned int cpu) |
2667 | { |
2668 | mce_intel_hcpu_update(cpu); |
2669 | |
2670 | /* intentionally ignoring frozen here */ |
2671 | if (!cpuhp_tasks_frozen) |
2672 | cmci_rediscover(); |
2673 | return 0; |
2674 | } |
2675 | |
2676 | static int mce_cpu_online(unsigned int cpu) |
2677 | { |
2678 | struct timer_list *t = this_cpu_ptr(&mce_timer); |
2679 | int ret; |
2680 | |
2681 | mce_device_create(cpu); |
2682 | |
2683 | ret = mce_threshold_create_device(cpu); |
2684 | if (ret) { |
2685 | mce_device_remove(cpu); |
2686 | return ret; |
2687 | } |
2688 | mce_reenable_cpu(); |
2689 | mce_start_timer(t); |
2690 | return 0; |
2691 | } |
2692 | |
2693 | static int mce_cpu_pre_down(unsigned int cpu) |
2694 | { |
2695 | struct timer_list *t = this_cpu_ptr(&mce_timer); |
2696 | |
2697 | mce_disable_cpu(); |
2698 | del_timer_sync(timer: t); |
2699 | mce_threshold_remove_device(cpu); |
2700 | mce_device_remove(cpu); |
2701 | return 0; |
2702 | } |
2703 | |
2704 | static __init void mce_init_banks(void) |
2705 | { |
2706 | int i; |
2707 | |
2708 | for (i = 0; i < MAX_NR_BANKS; i++) { |
2709 | struct mce_bank_dev *b = &mce_bank_devs[i]; |
2710 | struct device_attribute *a = &b->attr; |
2711 | |
2712 | b->bank = i; |
2713 | |
2714 | sysfs_attr_init(&a->attr); |
2715 | a->attr.name = b->attrname; |
2716 | snprintf(buf: b->attrname, ATTR_LEN, fmt: "bank%d" , i); |
2717 | |
2718 | a->attr.mode = 0644; |
2719 | a->show = show_bank; |
2720 | a->store = set_bank; |
2721 | } |
2722 | } |
2723 | |
2724 | /* |
2725 | * When running on XEN, this initcall is ordered against the XEN mcelog |
2726 | * initcall: |
2727 | * |
2728 | * device_initcall(xen_late_init_mcelog); |
2729 | * device_initcall_sync(mcheck_init_device); |
2730 | */ |
2731 | static __init int mcheck_init_device(void) |
2732 | { |
2733 | int err; |
2734 | |
2735 | /* |
2736 | * Check if we have a spare virtual bit. This will only become |
2737 | * a problem if/when we move beyond 5-level page tables. |
2738 | */ |
2739 | MAYBE_BUILD_BUG_ON(__VIRTUAL_MASK_SHIFT >= 63); |
2740 | |
2741 | if (!mce_available(c: &boot_cpu_data)) { |
2742 | err = -EIO; |
2743 | goto err_out; |
2744 | } |
2745 | |
2746 | if (!zalloc_cpumask_var(mask: &mce_device_initialized, GFP_KERNEL)) { |
2747 | err = -ENOMEM; |
2748 | goto err_out; |
2749 | } |
2750 | |
2751 | mce_init_banks(); |
2752 | |
2753 | err = subsys_system_register(subsys: &mce_subsys, NULL); |
2754 | if (err) |
2755 | goto err_out_mem; |
2756 | |
2757 | err = cpuhp_setup_state(state: CPUHP_X86_MCE_DEAD, name: "x86/mce:dead" , NULL, |
2758 | teardown: mce_cpu_dead); |
2759 | if (err) |
2760 | goto err_out_mem; |
2761 | |
2762 | /* |
2763 | * Invokes mce_cpu_online() on all CPUs which are online when |
2764 | * the state is installed. |
2765 | */ |
2766 | err = cpuhp_setup_state(state: CPUHP_AP_ONLINE_DYN, name: "x86/mce:online" , |
2767 | startup: mce_cpu_online, teardown: mce_cpu_pre_down); |
2768 | if (err < 0) |
2769 | goto err_out_online; |
2770 | |
2771 | register_syscore_ops(ops: &mce_syscore_ops); |
2772 | |
2773 | return 0; |
2774 | |
2775 | err_out_online: |
2776 | cpuhp_remove_state(state: CPUHP_X86_MCE_DEAD); |
2777 | |
2778 | err_out_mem: |
2779 | free_cpumask_var(mask: mce_device_initialized); |
2780 | |
2781 | err_out: |
2782 | pr_err("Unable to init MCE device (rc: %d)\n" , err); |
2783 | |
2784 | return err; |
2785 | } |
2786 | device_initcall_sync(mcheck_init_device); |
2787 | |
2788 | /* |
2789 | * Old style boot options parsing. Only for compatibility. |
2790 | */ |
2791 | static int __init mcheck_disable(char *str) |
2792 | { |
2793 | mca_cfg.disabled = 1; |
2794 | return 1; |
2795 | } |
2796 | __setup("nomce" , mcheck_disable); |
2797 | |
2798 | #ifdef CONFIG_DEBUG_FS |
2799 | struct dentry *mce_get_debugfs_dir(void) |
2800 | { |
2801 | static struct dentry *dmce; |
2802 | |
2803 | if (!dmce) |
2804 | dmce = debugfs_create_dir(name: "mce" , NULL); |
2805 | |
2806 | return dmce; |
2807 | } |
2808 | |
2809 | static void mce_reset(void) |
2810 | { |
2811 | atomic_set(v: &mce_fake_panicked, i: 0); |
2812 | atomic_set(v: &mce_executing, i: 0); |
2813 | atomic_set(v: &mce_callin, i: 0); |
2814 | atomic_set(v: &global_nwo, i: 0); |
2815 | cpumask_setall(dstp: &mce_missing_cpus); |
2816 | } |
2817 | |
2818 | static int fake_panic_get(void *data, u64 *val) |
2819 | { |
2820 | *val = fake_panic; |
2821 | return 0; |
2822 | } |
2823 | |
2824 | static int fake_panic_set(void *data, u64 val) |
2825 | { |
2826 | mce_reset(); |
2827 | fake_panic = val; |
2828 | return 0; |
2829 | } |
2830 | |
2831 | DEFINE_DEBUGFS_ATTRIBUTE(fake_panic_fops, fake_panic_get, fake_panic_set, |
2832 | "%llu\n" ); |
2833 | |
2834 | static void __init mcheck_debugfs_init(void) |
2835 | { |
2836 | struct dentry *dmce; |
2837 | |
2838 | dmce = mce_get_debugfs_dir(); |
2839 | debugfs_create_file_unsafe(name: "fake_panic" , mode: 0444, parent: dmce, NULL, |
2840 | fops: &fake_panic_fops); |
2841 | } |
2842 | #else |
2843 | static void __init mcheck_debugfs_init(void) { } |
2844 | #endif |
2845 | |
2846 | static int __init mcheck_late_init(void) |
2847 | { |
2848 | if (mca_cfg.recovery) |
2849 | enable_copy_mc_fragile(); |
2850 | |
2851 | mcheck_debugfs_init(); |
2852 | |
2853 | /* |
2854 | * Flush out everything that has been logged during early boot, now that |
2855 | * everything has been initialized (workqueues, decoders, ...). |
2856 | */ |
2857 | mce_schedule_work(); |
2858 | |
2859 | return 0; |
2860 | } |
2861 | late_initcall(mcheck_late_init); |
2862 | |