Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
Merge tag 'nmi.2023.02.14a' of git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu
Pull x86 NMI diagnostics from Paul McKenney:
"Add diagnostics to the x86 NMI handler to help detect NMI-handler bugs
on the one hand and failing hardware on the other"
* tag 'nmi.2023.02.14a' of git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu:
x86/nmi: Print reasons why backtrace NMIs are ignored
x86/nmi: Accumulate NMI-progress evidence in exc_nmi()
+128
-1
+107
-1
arch/x86/kernel/nmi.c
+107
-1
arch/x86/kernel/nmi.c
···
69
69
unsigned int unknown;
70
70
unsigned int external;
71
71
unsigned int swallow;
72
+
unsigned long recv_jiffies;
73
+
unsigned long idt_seq;
74
+
unsigned long idt_nmi_seq;
75
+
unsigned long idt_ignored;
76
+
atomic_long_t idt_calls;
77
+
unsigned long idt_seq_snap;
78
+
unsigned long idt_nmi_seq_snap;
79
+
unsigned long idt_ignored_snap;
80
+
long idt_calls_snap;
72
81
};
73
82
74
83
static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
···
488
479
DEFINE_IDTENTRY_RAW(exc_nmi)
489
480
{
490
481
irqentry_state_t irq_state;
482
+
struct nmi_stats *nsp = this_cpu_ptr(&nmi_stats);
491
483
492
484
/*
493
485
* Re-enable NMIs right here when running as an SEV-ES guest. This might
494
486
* cause nested NMIs, but those can be handled safely.
495
487
*/
496
488
sev_es_nmi_complete();
489
+
if (IS_ENABLED(CONFIG_NMI_CHECK_CPU))
490
+
arch_atomic_long_inc(&nsp->idt_calls);
497
491
498
492
if (IS_ENABLED(CONFIG_SMP) && arch_cpu_is_offline(smp_processor_id()))
499
493
return;
···
507
495
}
508
496
this_cpu_write(nmi_state, NMI_EXECUTING);
509
497
this_cpu_write(nmi_cr2, read_cr2());
498
+
if (IS_ENABLED(CONFIG_NMI_CHECK_CPU)) {
499
+
WRITE_ONCE(nsp->idt_seq, nsp->idt_seq + 1);
500
+
WARN_ON_ONCE(!(nsp->idt_seq & 0x1));
501
+
WRITE_ONCE(nsp->recv_jiffies, jiffies);
502
+
}
510
503
nmi_restart:
511
504
512
505
/*
···
526
509
527
510
inc_irq_stat(__nmi_count);
528
511
529
-
if (!ignore_nmis)
512
+
if (IS_ENABLED(CONFIG_NMI_CHECK_CPU) && ignore_nmis) {
513
+
WRITE_ONCE(nsp->idt_ignored, nsp->idt_ignored + 1);
514
+
} else if (!ignore_nmis) {
515
+
if (IS_ENABLED(CONFIG_NMI_CHECK_CPU)) {
516
+
WRITE_ONCE(nsp->idt_nmi_seq, nsp->idt_nmi_seq + 1);
517
+
WARN_ON_ONCE(!(nsp->idt_nmi_seq & 0x1));
518
+
}
530
519
default_do_nmi(regs);
520
+
if (IS_ENABLED(CONFIG_NMI_CHECK_CPU)) {
521
+
WRITE_ONCE(nsp->idt_nmi_seq, nsp->idt_nmi_seq + 1);
522
+
WARN_ON_ONCE(nsp->idt_nmi_seq & 0x1);
523
+
}
524
+
}
531
525
532
526
irqentry_nmi_exit(regs, irq_state);
533
527
···
553
525
554
526
if (user_mode(regs))
555
527
mds_user_clear_cpu_buffers();
528
+
if (IS_ENABLED(CONFIG_NMI_CHECK_CPU)) {
529
+
WRITE_ONCE(nsp->idt_seq, nsp->idt_seq + 1);
530
+
WARN_ON_ONCE(nsp->idt_seq & 0x1);
531
+
WRITE_ONCE(nsp->recv_jiffies, jiffies);
532
+
}
556
533
}
557
534
558
535
#if defined(CONFIG_X86_64) && IS_ENABLED(CONFIG_KVM_INTEL)
···
568
535
#endif
569
536
#if IS_MODULE(CONFIG_KVM_INTEL)
570
537
EXPORT_SYMBOL_GPL(asm_exc_nmi_noist);
538
+
#endif
539
+
540
+
#ifdef CONFIG_NMI_CHECK_CPU
541
+
542
+
static char *nmi_check_stall_msg[] = {
543
+
/* */
544
+
/* +--------- nsp->idt_seq_snap & 0x1: CPU is in NMI handler. */
545
+
/* | +------ cpu_is_offline(cpu) */
546
+
/* | | +--- nsp->idt_calls_snap != atomic_long_read(&nsp->idt_calls): */
547
+
/* | | | NMI handler has been invoked. */
548
+
/* | | | */
549
+
/* V V V */
550
+
/* 0 0 0 */ "NMIs are not reaching exc_nmi() handler",
551
+
/* 0 0 1 */ "exc_nmi() handler is ignoring NMIs",
552
+
/* 0 1 0 */ "CPU is offline and NMIs are not reaching exc_nmi() handler",
553
+
/* 0 1 1 */ "CPU is offline and exc_nmi() handler is legitimately ignoring NMIs",
554
+
/* 1 0 0 */ "CPU is in exc_nmi() handler and no further NMIs are reaching handler",
555
+
/* 1 0 1 */ "CPU is in exc_nmi() handler which is legitimately ignoring NMIs",
556
+
/* 1 1 0 */ "CPU is offline in exc_nmi() handler and no more NMIs are reaching exc_nmi() handler",
557
+
/* 1 1 1 */ "CPU is offline in exc_nmi() handler which is legitimately ignoring NMIs",
558
+
};
559
+
560
+
void nmi_backtrace_stall_snap(const struct cpumask *btp)
561
+
{
562
+
int cpu;
563
+
struct nmi_stats *nsp;
564
+
565
+
for_each_cpu(cpu, btp) {
566
+
nsp = per_cpu_ptr(&nmi_stats, cpu);
567
+
nsp->idt_seq_snap = READ_ONCE(nsp->idt_seq);
568
+
nsp->idt_nmi_seq_snap = READ_ONCE(nsp->idt_nmi_seq);
569
+
nsp->idt_ignored_snap = READ_ONCE(nsp->idt_ignored);
570
+
nsp->idt_calls_snap = atomic_long_read(&nsp->idt_calls);
571
+
}
572
+
}
573
+
574
+
void nmi_backtrace_stall_check(const struct cpumask *btp)
575
+
{
576
+
int cpu;
577
+
int idx;
578
+
unsigned long nmi_seq;
579
+
unsigned long j = jiffies;
580
+
char *modp;
581
+
char *msgp;
582
+
char *msghp;
583
+
struct nmi_stats *nsp;
584
+
585
+
for_each_cpu(cpu, btp) {
586
+
nsp = per_cpu_ptr(&nmi_stats, cpu);
587
+
modp = "";
588
+
msghp = "";
589
+
nmi_seq = READ_ONCE(nsp->idt_nmi_seq);
590
+
if (nsp->idt_nmi_seq_snap + 1 == nmi_seq && (nmi_seq & 0x1)) {
591
+
msgp = "CPU entered NMI handler function, but has not exited";
592
+
} else if ((nsp->idt_nmi_seq_snap & 0x1) != (nmi_seq & 0x1)) {
593
+
msgp = "CPU is handling NMIs";
594
+
} else {
595
+
idx = ((nsp->idt_seq_snap & 0x1) << 2) |
596
+
(cpu_is_offline(cpu) << 1) |
597
+
(nsp->idt_calls_snap != atomic_long_read(&nsp->idt_calls));
598
+
msgp = nmi_check_stall_msg[idx];
599
+
if (nsp->idt_ignored_snap != READ_ONCE(nsp->idt_ignored) && (idx & 0x1))
600
+
modp = ", but OK because ignore_nmis was set";
601
+
if (nmi_seq & ~0x1)
602
+
msghp = " (CPU currently in NMI handler function)";
603
+
else if (nsp->idt_nmi_seq_snap + 1 == nmi_seq)
604
+
msghp = " (CPU exited one NMI handler function)";
605
+
}
606
+
pr_alert("%s: CPU %d: %s%s%s, last activity: %lu jiffies ago.\n",
607
+
__func__, cpu, msgp, modp, msghp, j - READ_ONCE(nsp->recv_jiffies));
608
+
}
609
+
}
610
+
571
611
#endif
572
612
573
613
void stop_nmi(void)
+8
include/linux/nmi.h
+8
include/linux/nmi.h
···
214
214
#include <asm/nmi.h>
215
215
#endif
216
216
217
+
#ifdef CONFIG_NMI_CHECK_CPU
218
+
void nmi_backtrace_stall_snap(const struct cpumask *btp);
219
+
void nmi_backtrace_stall_check(const struct cpumask *btp);
220
+
#else
221
+
static inline void nmi_backtrace_stall_snap(const struct cpumask *btp) {}
222
+
static inline void nmi_backtrace_stall_check(const struct cpumask *btp) {}
223
+
#endif
224
+
217
225
#endif
+11
lib/Kconfig.debug
+11
lib/Kconfig.debug
···
1562
1562
depends on TRACE_IRQFLAGS
1563
1563
depends on TRACE_IRQFLAGS_NMI_SUPPORT
1564
1564
1565
+
config NMI_CHECK_CPU
1566
+
bool "Debugging for CPUs failing to respond to backtrace requests"
1567
+
depends on DEBUG_KERNEL
1568
+
depends on X86
1569
+
default n
1570
+
help
1571
+
Enables debug prints when a CPU fails to respond to a given
1572
+
backtrace NMI. These prints provide some reasons why a CPU
1573
+
might legitimately be failing to respond, for example, if it
1574
+
is offline of if ignore_nmis is set.
1575
+
1565
1576
config DEBUG_IRQFLAGS
1566
1577
bool "Debug IRQ flag manipulation"
1567
1578
help
+2
lib/nmi_backtrace.c
+2
lib/nmi_backtrace.c
···
64
64
if (!cpumask_empty(to_cpumask(backtrace_mask))) {
65
65
pr_info("Sending NMI from CPU %d to CPUs %*pbl:\n",
66
66
this_cpu, nr_cpumask_bits, to_cpumask(backtrace_mask));
67
+
nmi_backtrace_stall_snap(to_cpumask(backtrace_mask));
67
68
raise(to_cpumask(backtrace_mask));
68
69
}
69
70
···
75
74
mdelay(1);
76
75
touch_softlockup_watchdog();
77
76
}
77
+
nmi_backtrace_stall_check(to_cpumask(backtrace_mask));
78
78
79
79
/*
80
80
* Force flush any remote buffers that might be stuck in IRQ context