Skip to contentSkip to conversation

x86, nmi: new NMI handling routines

PATCH Superseded Don Zickus 6 participants 28 messages Sep 20, 2011 – Sep 25, 2011

Subsystems (8)

v5 of 7 Superseded Latest is v7 →

lore
Share & tools
lore

Don Zickus

author
#0 lore raw

v5 cover x86, nmi: new NMI handling routines

I had the pleasure of hosting Robert Richter at Red Hat in August. One of the issues he wanted to talk with me about was having the NMI handling routines execute all the NMI handlers for each NMI mainly for his AMD IBS work. But he also brought up another good point that because of the way NMIs work, it is possible to lose them if multiple NMIs happen at the same time.

As a result, we sat around and discussed how we could go about executing all the nmi handlers for each NMI to ensure that we would not lose any events.

We decided the best way to do this would be to have the NMI handlers break away from the notifier routines and create our own. This would allow us to execute all the handlers without hacking up the notifier stuff and easily track the number of events processed at a higher level to deal with the new problemm of extra NMIs.

I spent some time hacking and came up with this patch. I tested it on my core2quad machine trying to enable all the NMI handler I could, mainly perf and kgdb (and oprofile too when perf was disabled). Everything seems to work correctly. If people are ok with this approach, I'll try and test this on more machines.

More details about the patch are in the individual changelogs.

This version only changes patch 4.

Don Zickus (6):
x86, nmi: split out nmi from traps.c
x86, nmi: create new NMI handler routines
x86, nmi: wire up NMI handlers to new routines
x86, nmi: add in logic to handle multiple events and unknown NMIs x86, nmi: track NMI usage stats
x86, nmi: print out NMI stats in /proc/interrupts

 arch/x86/include/asm/nmi.h              |   39 ++--
 arch/x86/include/asm/reboot.h           |    2 +-
 arch/x86/kernel/Makefile                |    2 +-
 arch/x86/kernel/apic/hw_nmi.c           |   27 +--
 arch/x86/kernel/apic/x2apic_uv_x.c      |   20 +-
 arch/x86/kernel/cpu/mcheck/mce-inject.c |   20 +-
 arch/x86/kernel/cpu/mcheck/mce.c        |    3 -
 arch/x86/kernel/cpu/perf_event.c        |   60 +----
 arch/x86/kernel/crash.c                 |    5 +-
 arch/x86/kernel/irq.c                   |    2 +
 arch/x86/kernel/kgdb.c                  |   60 +++-
 arch/x86/kernel/nmi.c                   |  473 +++++++++++++++++++++++++++++++
 arch/x86/kernel/reboot.c                |   23 +-
 arch/x86/kernel/traps.c                 |  155 ----------
 arch/x86/oprofile/nmi_int.c             |   40 +--
 arch/x86/oprofile/nmi_timer_int.c       |   28 +--
 drivers/acpi/apei/ghes.c                |   22 +-
 drivers/char/ipmi/ipmi_watchdog.c       |   32 +--
 drivers/watchdog/hpwdt.c                |   23 +--
 19 files changed, 613 insertions(+), 423 deletions(-)
 create mode 100644 arch/x86/kernel/nmi.c
Signature · 1 line
1.7.6
author
#1 lore raw

PATCH v5 1/6 x86, nmi: split out nmi from traps.c

The nmi stuff is changing a lot and adding more functionality. Split it out from the traps.c file so it doesn't continue to pollute that file.

This makes it easier to find and expand all the future nmi related work.

No real functional changes here.

---
 arch/x86/kernel/Makefile |    2 +-
 arch/x86/kernel/nmi.c    |  178 ++++++++++++++++++++++++++++++++++++++++++++++
 arch/x86/kernel/traps.c  |  155 ----------------------------------------
 3 files changed, 179 insertions(+), 156 deletions(-)
 create mode 100644 arch/x86/kernel/nmi.c
arch/x86/kernel/Makefile +1 -1
index 82f2912..8baca3c 100644
@@ -19,7 +19,7 @@ endif
 
 obj-y			:= process_$(BITS).o signal.o entry_$(BITS).o
 obj-y			+= traps.o irq.o irq_$(BITS).o dumpstack_$(BITS).o
-obj-y			+= time.o ioport.o ldt.o dumpstack.o
+obj-y			+= time.o ioport.o ldt.o dumpstack.o nmi.o
 obj-y			+= setup.o x86_init.o i8259.o irqinit.o jump_label.o
 obj-$(CONFIG_IRQ_WORK)  += irq_work.o
 obj-y			+= probe_roms.o
arch/x86/kernel/nmi.c new file +178 -0

178 lines not on this page · raw message

arch/x86/kernel/traps.c +0 -155

167 lines not on this page · raw message

Signature · 1 line
1.7.6
author
#2 lore raw

PATCH v5 2/6 x86, nmi: create new NMI handler routines

In reply to Don Zickus

The NMI handlers used to rely on the notifier infrastructure. This worked great until we wanted to support handling multiple events better.

One of the key ideas to the nmi handling is to process _all_ the handlers for each NMI. The reason behind this switch is because NMIs are edge triggered. If enough NMIs are triggered, then they could be lost because the cpu can only latch at most one NMI (besides the one currently being processed).

In order to deal with this we have decided to process all the NMI handlers for each NMI. This allows the handlers to determine if they recieved an event or not (the ones that can not determine this will be left to fend for themselves on the unknown NMI list).

As a result of this change it is now possible to have an extra NMI that was destined to be received for an already processed event. Because the event was processed in the previous NMI, this NMI gets dropped and becomes an 'unknown' NMI. This of course will cause printks that scare people.

However, we prefer to have extra NMIs as opposed to losing NMIs and as such are have developed a basic mechanism to catch most of them. That will be a later patch.

To accomplish this idea, I unhooked the nmi handlers from the notifier routines and created a new mechanism loosely based on doIRQ. The reason for this is the notifier routines have a couple of shortcomings. One we could't guarantee all future NMI handlers used NOTIFY_OK instead of NOTIFY_STOP. Second, we couldn't keep track of the number of events being handled in each routine (most only handle one, perf can handle more than one). Third, I wanted to eventually display which nmi handlers are registered in the system in /proc/interrupts to help see who is generating NMIs.

The patch below just implements the new infrastructure but doesn't wire it up yet (that is the next patch). Its design is based on doIRQ structs and the atomic notifier routines. So the rcu stuff in the patch isn't entirely untested (as the notifier routines have soaked it) but it should be double checked in case I copied the code wrong.

V2:
- use kstrdup to copy/allocate device name
- fix-up _GPL oops

V3:
- fix leak in register_nmi_handler error path
- removed _raw annotations from rcu_dereference

V4:
- handle kstrndup failure

---
 arch/x86/include/asm/nmi.h |   19 +++++
 arch/x86/kernel/nmi.c      |  157 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 176 insertions(+), 0 deletions(-)
arch/x86/include/asm/nmi.h +19 -0
index 4886a68..6d04b28 100644
@@ -42,6 +42,25 @@ void arch_trigger_all_cpu_backtrace(void);
 #define NMI_LOCAL_NORMAL_PRIOR	(NMI_LOCAL_BIT | NMI_NORMAL_PRIOR)
 #define NMI_LOCAL_LOW_PRIOR	(NMI_LOCAL_BIT | NMI_LOW_PRIOR)
 
+#define NMI_FLAG_FIRST	1
+
+enum {
+	NMI_LOCAL=0,
+	NMI_UNKNOWN,
+	NMI_EXTERNAL,
+	NMI_MAX
+};
+
+#define NMI_DONE	0
+#define NMI_HANDLED	1
+
+typedef int (*nmi_handler_t)(unsigned int, struct pt_regs *);
+
+int register_nmi_handler(unsigned int, nmi_handler_t, unsigned long,
+			 const char *);
+
+void unregister_nmi_handler(unsigned int, const char *);
+
 void stop_nmi(void);
 void restart_nmi(void);
 
arch/x86/kernel/nmi.c +157 -0

175 lines not on this page · raw message

Signature · 1 line
1.7.6
#3 lore raw

Re: [V5][PATCH 2/6] x86, nmi: create new NMI handler routines

Quoted 99 lines · Don Zickus

The NMI handlers used to rely on the notifier infrastructure. This worked great until we wanted to support handling multiple events better.

One of the key ideas to the nmi handling is to process _all_ the handlers for each NMI. The reason behind this switch is because NMIs are edge triggered. If enough NMIs are triggered, then they could be lost because the cpu can only latch at most one NMI (besides the one currently being processed).

In order to deal with this we have decided to process all the NMI handlers for each NMI. This allows the handlers to determine if they recieved an event or not (the ones that can not determine this will be left to fend for themselves on the unknown NMI list).

As a result of this change it is now possible to have an extra NMI that was destined to be received for an already processed event. Because the event was processed in the previous NMI, this NMI gets dropped and becomes an 'unknown' NMI. This of course will cause printks that scare people.

However, we prefer to have extra NMIs as opposed to losing NMIs and as such are have developed a basic mechanism to catch most of them. That will be a later patch.

To accomplish this idea, I unhooked the nmi handlers from the notifier routines and created a new mechanism loosely based on doIRQ. The reason for this is the notifier routines have a couple of shortcomings. One we could't guarantee all future NMI handlers used NOTIFY_OK instead of NOTIFY_STOP. Second, we couldn't keep track of the number of events being handled in each routine (most only handle one, perf can handle more than one). Third, I wanted to eventually display which nmi handlers are registered in the system in /proc/interrupts to help see who is generating NMIs.

The patch below just implements the new infrastructure but doesn't wire it up yet (that is the next patch). Its design is based on doIRQ structs and the atomic notifier routines. So the rcu stuff in the patch isn't entirely untested (as the notifier routines have soaked it) but it should be double checked in case I copied the code wrong.

V2:
- use kstrdup to copy/allocate device name
- fix-up _GPL oops

V3:
- fix leak in register_nmi_handler error path
- removed _raw annotations from rcu_dereference

V4:
- handle kstrndup failure

---
 arch/x86/include/asm/nmi.h |   19 +++++
 arch/x86/kernel/nmi.c      |  157 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 176 insertions(+), 0 deletions(-)
arch/x86/include/asm/nmi.h +19 -0
index 4886a68..6d04b28 100644
@@ -42,6 +42,25 @@ void arch_trigger_all_cpu_backtrace(void);
 #define NMI_LOCAL_NORMAL_PRIOR	(NMI_LOCAL_BIT | NMI_NORMAL_PRIOR)
 #define NMI_LOCAL_LOW_PRIOR	(NMI_LOCAL_BIT | NMI_LOW_PRIOR)
 
+#define NMI_FLAG_FIRST	1
+
+enum {
+	NMI_LOCAL=0,
+	NMI_UNKNOWN,
+	NMI_EXTERNAL,
+	NMI_MAX
+};
+
+#define NMI_DONE	0
+#define NMI_HANDLED	1
+
+typedef int (*nmi_handler_t)(unsigned int, struct pt_regs *);
+
+int register_nmi_handler(unsigned int, nmi_handler_t, unsigned long,
+			 const char *);
+
+void unregister_nmi_handler(unsigned int, const char *);
+
 void stop_nmi(void);
 void restart_nmi(void);
 
arch/x86/kernel/nmi.c +3 -0
index 68d758a..c2df58a 100644
@@ -13,6 +13,9 @@
 #include <linux/kprobes.h>
 #include <linux/kdebug.h>
 #include <linux/nmi.h>
+#include <linux/delay.h>
+#include <linux/hardirq.h>
+#include <linux/slab.h>
 
 #if defined(CONFIG_EDAC)
 #include <linux/edac.h>

arch/x86/kernel/nmi.c line 31

Quoted 8 lines
arch/x86/kernel/nmi.c +5 -0
@@ -21,6 +24,28 @@
 #include <linux/atomic.h>
 #include <asm/traps.h>
 #include <asm/mach_traps.h>
+#include <asm/nmi.h>
+
+#define NMI_MAX_NAMELEN	16
+struct nmiaction {
+	struct nmiaction __rcu *next;

Why not just use struct list_head here and use list_xxx_rcu family to operate on the list? IMHO, that will make code simpler without much overhead.

Best Regards,
Huang Ying

author
#4 lore raw

Re: [V5][PATCH 2/6] x86, nmi: create new NMI handler routines

Quoted 12 lines · Huang Ying
Quoted 11 lines · Don Zickus

index 68d758a..c2df58a 100644

arch/x86/kernel/nmi.c +3 -0
@@ -13,6 +13,9 @@
 #include <linux/kprobes.h>
 #include <linux/kdebug.h>
 #include <linux/nmi.h>
+#include <linux/delay.h>
+#include <linux/hardirq.h>
+#include <linux/slab.h>
 
Quoted 15 lines

arch/x86/kernel/nmi.c line 31

Quoted 11 lines
arch/x86/kernel/nmi.c +5 -0
 #if defined(CONFIG_EDAC)
 #include <linux/edac.h>
@@ -21,6 +24,28 @@
 #include <linux/atomic.h>
 #include <asm/traps.h>
 #include <asm/mach_traps.h>
+#include <asm/nmi.h>
+
+#define NMI_MAX_NAMELEN	16
+struct nmiaction {
+	struct nmiaction __rcu *next;

Why not just use struct list_head here and use list_xxx_rcu family to operate on the list? IMHO, that will make code simpler without much overhead.

Yeah, I just copied and pasted from what was in the notifier chain stuff. Even though I don't need the previous pointer, the awkward rcu pointer stuff would be a lot simpler to implement and read with those macros. Let me play with it and see what I come up with.

Thanks for the suggestion.

Cheers,
Don

author
#5 lore raw

PATCH v5 3/6 x86, nmi: wire up NMI handlers to new routines

In reply to Don Zickus

Just convert all the files that have an nmi handler to the new routines. Most of it is straight forward conversion. A couple of places needed some tweaking like kgdb which separates the debug notifier from the nmi handler and mce removes a call to notify_die (as I couldn't figure out why it was there).

The things that get converted are the registeration/unregistration routines and the nmi handler itself has its args changed along with code removal to check which list it is on (most are on one NMI list except for kgdb which has both an NMI routine and an NMI Unknown routine).

Cc Jason Wessel <[email protected]>
Cc Andi Kleen <[email protected]>
Cc Robert Richter <[email protected]>
Cc Huang Ying <[email protected]>
Cc Corey Minyard <[email protected]>
Cc Jack Steiner <[email protected]>
---
 arch/x86/include/asm/nmi.h              |   20 ----------
 arch/x86/include/asm/reboot.h           |    2 +-
 arch/x86/kernel/apic/hw_nmi.c           |   27 +++-----------
 arch/x86/kernel/apic/x2apic_uv_x.c      |   20 ++--------
 arch/x86/kernel/cpu/mcheck/mce-inject.c |   20 ++++-------
 arch/x86/kernel/cpu/mcheck/mce.c        |    3 --
 arch/x86/kernel/cpu/perf_event.c        |   60 +++----------------------------
 arch/x86/kernel/crash.c                 |    5 +--
 arch/x86/kernel/kgdb.c                  |   60 +++++++++++++++++++++++--------
 arch/x86/kernel/nmi.c                   |   11 ++++--
 arch/x86/kernel/reboot.c                |   23 ++++--------
 arch/x86/oprofile/nmi_int.c             |   40 ++++++--------------
 arch/x86/oprofile/nmi_timer_int.c       |   28 +++-----------
 drivers/acpi/apei/ghes.c                |   22 ++++-------
 drivers/char/ipmi/ipmi_watchdog.c       |   32 +++++-----------
 drivers/watchdog/hpwdt.c                |   23 +++---------
 16 files changed, 125 insertions(+), 271 deletions(-)
arch/x86/include/asm/nmi.h +0 -20
index 6d04b28..fc74547 100644
@@ -22,26 +22,6 @@ void arch_trigger_all_cpu_backtrace(void);
 #define arch_trigger_all_cpu_backtrace arch_trigger_all_cpu_backtrace
 #endif
 
-/*
- * Define some priorities for the nmi notifier call chain.
- *
- * Create a local nmi bit that has a higher priority than
- * external nmis, because the local ones are more frequent.
- *
- * Also setup some default high/normal/low settings for
- * subsystems to registers with.  Using 4 bits to separate
- * the priorities.  This can go a lot higher if needed be.
- */
-
-#define NMI_LOCAL_SHIFT		16	/* randomly picked */
-#define NMI_LOCAL_BIT		(1ULL << NMI_LOCAL_SHIFT)
-#define NMI_HIGH_PRIOR		(1ULL << 8)
-#define NMI_NORMAL_PRIOR	(1ULL << 4)
-#define NMI_LOW_PRIOR		(1ULL << 0)
-#define NMI_LOCAL_HIGH_PRIOR	(NMI_LOCAL_BIT | NMI_HIGH_PRIOR)
-#define NMI_LOCAL_NORMAL_PRIOR	(NMI_LOCAL_BIT | NMI_NORMAL_PRIOR)
-#define NMI_LOCAL_LOW_PRIOR	(NMI_LOCAL_BIT | NMI_LOW_PRIOR)
-
 #define NMI_FLAG_FIRST	1
 
 enum {
arch/x86/include/asm/reboot.h +1 -1
index 3250e3d..92f29706 100644
@@ -23,7 +23,7 @@ void machine_real_restart(unsigned int type);
 #define MRR_BIOS	0
 #define MRR_APM		1
 
-typedef void (*nmi_shootdown_cb)(int, struct die_args*);
+typedef void (*nmi_shootdown_cb)(int, struct pt_regs*);
 void nmi_shootdown_cpus(nmi_shootdown_cb callback);
 
 #endif /* _ASM_X86_REBOOT_H */
arch/x86/kernel/apic/hw_nmi.c +5 -22
index d5e57db0..31cb9ae 100644
@@ -60,22 +60,10 @@ void arch_trigger_all_cpu_backtrace(void)
 }
 
 static int __kprobes
-arch_trigger_all_cpu_backtrace_handler(struct notifier_block *self,
-			 unsigned long cmd, void *__args)
+arch_trigger_all_cpu_backtrace_handler(unsigned int cmd, struct pt_regs *regs)
 {
-	struct die_args *args = __args;
-	struct pt_regs *regs;
 	int cpu;
 
-	switch (cmd) {
-	case DIE_NMI:
-		break;
-
-	default:
-		return NOTIFY_DONE;
-	}
-
-	regs = args->regs;
 	cpu = smp_processor_id();
 
 	if (cpumask_test_cpu(cpu, to_cpumask(backtrace_mask))) {
@@ -86,21 +74,16 @@ arch_trigger_all_cpu_backtrace_handler(struct notifier_block *self,
 		show_regs(regs);
 		arch_spin_unlock(&lock);
 		cpumask_clear_cpu(cpu, to_cpumask(backtrace_mask));
-		return NOTIFY_STOP;
+		return NMI_HANDLED;
 	}
 
-	return NOTIFY_DONE;
+	return NMI_DONE;
 }
 
-static __read_mostly struct notifier_block backtrace_notifier = {
-	.notifier_call          = arch_trigger_all_cpu_backtrace_handler,
-	.next                   = NULL,
-	.priority               = NMI_LOCAL_LOW_PRIOR,
-};
-
 static int __init register_trigger_all_cpu_backtrace(void)
 {
-	register_die_notifier(&backtrace_notifier);
+	register_nmi_handler(NMI_LOCAL, arch_trigger_all_cpu_backtrace_handler,
+				0, "arch_bt");
 	return 0;
 }
 early_initcall(register_trigger_all_cpu_backtrace);
arch/x86/kernel/apic/x2apic_uv_x.c +4 -16
index 34b1859..75be00e 100644
@@ -672,18 +672,11 @@ void __cpuinit uv_cpu_init(void)
 /*
  * When NMI is received, print a stack trace.
  */
-int uv_handle_nmi(struct notifier_block *self, unsigned long reason, void *data)
+int uv_handle_nmi(unsigned int reason, struct pt_regs *regs)
 {
 	unsigned long real_uv_nmi;
 	int bid;
 
-	if (reason != DIE_NMIUNKNOWN)
-		return NOTIFY_OK;
-
-	if (in_crash_kexec)
-		/* do nothing if entering the crash kernel */
-		return NOTIFY_OK;
-
 	/*
 	 * Each blade has an MMR that indicates when an NMI has been sent
 	 * to cpus on the blade. If an NMI is detected, atomically
@@ -704,7 +697,7 @@ int uv_handle_nmi(struct notifier_block *self, unsigned long reason, void *data)
 	}
 
 	if (likely(__get_cpu_var(cpu_last_nmi_count) == uv_blade_info[bid].nmi_count))
-		return NOTIFY_DONE;
+		return NMI_DONE;
 
 	__get_cpu_var(cpu_last_nmi_count) = uv_blade_info[bid].nmi_count;
 
@@ -717,17 +710,12 @@ int uv_handle_nmi(struct notifier_block *self, unsigned long reason, void *data)
 	dump_stack();
 	spin_unlock(&uv_nmi_lock);
 
-	return NOTIFY_STOP;
+	return NMI_HANDLED;
 }
 
-static struct notifier_block uv_dump_stack_nmi_nb = {
-	.notifier_call	= uv_handle_nmi,
-	.priority = NMI_LOCAL_LOW_PRIOR - 1,
-};
-
 void uv_register_nmi_notifier(void)
 {
-	if (register_die_notifier(&uv_dump_stack_nmi_nb))
+	if (register_nmi_handler(NMI_UNKNOWN, uv_handle_nmi, 0, "uv"))
 		printk(KERN_WARNING "UV NMI handler failed to register\n");
 }
 
arch/x86/kernel/cpu/mcheck/mce-inject.c +7 -13
index 0ed633c..6199232 100644
@@ -78,27 +78,20 @@ static void raise_exception(struct mce *m, struct pt_regs *pregs)
 
 static cpumask_var_t mce_inject_cpumask;
 
-static int mce_raise_notify(struct notifier_block *self,
-			    unsigned long val, void *data)
+static int mce_raise_notify(unsigned int cmd, struct pt_regs *regs)
 {
-	struct die_args *args = (struct die_args *)data;
 	int cpu = smp_processor_id();
 	struct mce *m = &__get_cpu_var(injectm);
-	if (val != DIE_NMI || !cpumask_test_cpu(cpu, mce_inject_cpumask))
-		return NOTIFY_DONE;
+	if (!cpumask_test_cpu(cpu, mce_inject_cpumask))
+		return NMI_DONE;
 	cpumask_clear_cpu(cpu, mce_inject_cpumask);
 	if (m->inject_flags & MCJ_EXCEPTION)
-		raise_exception(m, args->regs);
+		raise_exception(m, regs);
 	else if (m->status)
 		raise_poll(m);
-	return NOTIFY_STOP;
+	return NMI_HANDLED;
 }
 
-static struct notifier_block mce_raise_nb = {
-	.notifier_call = mce_raise_notify,
-	.priority = NMI_LOCAL_NORMAL_PRIOR,
-};
-
 /* Inject mce on current CPU */
 static int raise_local(void)
 {
@@ -216,7 +209,8 @@ static int inject_init(void)
 		return -ENOMEM;
 	printk(KERN_INFO "Machine check injector initialized\n");
 	mce_chrdev_ops.write = mce_write;
-	register_die_notifier(&mce_raise_nb);
+	register_nmi_handler(NMI_LOCAL, mce_raise_notify, 0,
+				"mce_notify");
 	return 0;
 }
 
arch/x86/kernel/cpu/mcheck/mce.c +0 -3
index 08363b0..3fc65b6 100644
@@ -908,9 +908,6 @@ void do_machine_check(struct pt_regs *regs, long error_code)
 
 	percpu_inc(mce_exception_count);
 
-	if (notify_die(DIE_NMI, "machine check", regs, error_code,
-			   18, SIGKILL) == NOTIFY_STOP)
-		goto out;
 	if (!banks)
 		goto out;
 
arch/x86/kernel/cpu/perf_event.c +5 -55

80 lines not on this page · raw message

arch/x86/kernel/crash.c +1 -4
index 764c7c2..13ad899 100644
@@ -32,15 +32,12 @@ int in_crash_kexec;
 
 #if defined(CONFIG_SMP) && defined(CONFIG_X86_LOCAL_APIC)
 
-static void kdump_nmi_callback(int cpu, struct die_args *args)
+static void kdump_nmi_callback(int cpu, struct pt_regs *regs)
 {
-	struct pt_regs *regs;
 #ifdef CONFIG_X86_32
 	struct pt_regs fixed_regs;
 #endif
 
-	regs = args->regs;
-
 #ifdef CONFIG_X86_32
 	if (!user_mode_vm(regs)) {
 		crash_fixup_ss_esp(&fixed_regs, regs);
arch/x86/kernel/kgdb.c +45 -15

97 lines not on this page · raw message

arch/x86/kernel/nmi.c +8 -3
index c2df58a..acd61e8 100644
@@ -1,6 +1,7 @@
 /*
  *  Copyright (C) 1991, 1992  Linus Torvalds
  *  Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs
+ *  Copyright (C) 2011	Don Zickus Red Hat, Inc.
  *
  *  Pentium III FXSR, SSE support
  *	Gareth Hughes <[email protected]>, May 2000
@@ -252,8 +253,10 @@ io_check_error(unsigned char reason, struct pt_regs *regs)
 static notrace __kprobes void
 unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
 {
-	if (notify_die(DIE_NMIUNKNOWN, "nmi", regs, reason, 2, SIGINT) ==
-			NOTIFY_STOP)
+	int handled;
+
+	handled = nmi_handle(NMI_UNKNOWN, regs);
+	if (handled) 
 		return;
 #ifdef CONFIG_MCA
 	/*
@@ -278,13 +281,15 @@ unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
 static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
 {
 	unsigned char reason = 0;
+	int handled;
 
 	/*
 	 * CPU-specific NMI must be processed before non-CPU-specific
 	 * NMI, otherwise we may lose it, because the CPU-specific
 	 * NMI can not be detected/processed on other CPUs.
 	 */
-	if (notify_die(DIE_NMI, "nmi", regs, 0, 2, SIGINT) == NOTIFY_STOP)
+	handled = nmi_handle(NMI_LOCAL, regs);
+	if (handled)
 		return;
 
 	/* Non-CPU-specific NMI: NMI sources can be processed on any CPU */
arch/x86/kernel/reboot.c +7 -16
index 9242436..adab340 100644
@@ -464,7 +464,7 @@ static inline void kb_wait(void)
 	}
 }
 
-static void vmxoff_nmi(int cpu, struct die_args *args)
+static void vmxoff_nmi(int cpu, struct pt_regs *regs)
 {
 	cpu_emergency_vmxoff();
 }
@@ -736,14 +736,10 @@ static nmi_shootdown_cb shootdown_callback;
 
 static atomic_t waiting_for_crash_ipi;
 
-static int crash_nmi_callback(struct notifier_block *self,
-			unsigned long val, void *data)
+static int crash_nmi_callback(unsigned int val, struct pt_regs *regs)
 {
 	int cpu;
 
-	if (val != DIE_NMI)
-		return NOTIFY_OK;
-
 	cpu = raw_smp_processor_id();
 
 	/* Don't do anything if this handler is invoked on crashing cpu.
@@ -751,10 +747,10 @@ static int crash_nmi_callback(struct notifier_block *self,
 	 * an NMI if system was initially booted with nmi_watchdog parameter.
 	 */
 	if (cpu == crashing_cpu)
-		return NOTIFY_STOP;
+		return NMI_HANDLED;
 	local_irq_disable();
 
-	shootdown_callback(cpu, (struct die_args *)data);
+	shootdown_callback(cpu, regs);
 
 	atomic_dec(&waiting_for_crash_ipi);
 	/* Assume hlt works */
@@ -762,7 +758,7 @@ static int crash_nmi_callback(struct notifier_block *self,
 	for (;;)
 		cpu_relax();
 
-	return 1;
+	return NMI_HANDLED;
 }
 
 static void smp_send_nmi_allbutself(void)
@@ -770,12 +766,6 @@ static void smp_send_nmi_allbutself(void)
 	apic->send_IPI_allbutself(NMI_VECTOR);
 }
 
-static struct notifier_block crash_nmi_nb = {
-	.notifier_call = crash_nmi_callback,
-	/* we want to be the first one called */
-	.priority = NMI_LOCAL_HIGH_PRIOR+1,
-};
-
 /* Halt all other CPUs, calling the specified function on each of them
  *
  * This function can be used to halt all other CPUs on crash
@@ -794,7 +784,8 @@ void nmi_shootdown_cpus(nmi_shootdown_cb callback)
 
 	atomic_set(&waiting_for_crash_ipi, num_online_cpus() - 1);
 	/* Would it be better to replace the trap vector here? */
-	if (register_die_notifier(&crash_nmi_nb))
+	if (register_nmi_handler(NMI_LOCAL, crash_nmi_callback, 
+				 NMI_FLAG_FIRST, "crash"))
 		return;		/* return what? */
 	/* Ensure the new callback function is set before sending
 	 * out the NMI
arch/x86/oprofile/nmi_int.c +11 -27
index 68894fd..adf8fb3 100644
@@ -61,26 +61,15 @@ u64 op_x86_get_ctrl(struct op_x86_model_spec const *model,
 }
 
 
-static int profile_exceptions_notify(struct notifier_block *self,
-				     unsigned long val, void *data)
+static int profile_exceptions_notify(unsigned int val, struct pt_regs *regs)
 {
-	struct die_args *args = (struct die_args *)data;
-	int ret = NOTIFY_DONE;
-
-	switch (val) {
-	case DIE_NMI:
-		if (ctr_running)
-			model->check_ctrs(args->regs, &__get_cpu_var(cpu_msrs));
-		else if (!nmi_enabled)
-			break;
-		else
-			model->stop(&__get_cpu_var(cpu_msrs));
-		ret = NOTIFY_STOP;
-		break;
-	default:
-		break;
-	}
-	return ret;
+	if (ctr_running)
+		model->check_ctrs(regs, &__get_cpu_var(cpu_msrs));
+	else if (!nmi_enabled)
+		return NMI_DONE;
+	else
+		model->stop(&__get_cpu_var(cpu_msrs));
+	return NMI_HANDLED;
 }
 
 static void nmi_cpu_save_registers(struct op_msrs *msrs)
@@ -363,12 +352,6 @@ static void nmi_cpu_setup(void *dummy)
 	apic_write(APIC_LVTPC, APIC_DM_NMI);
 }
 
-static struct notifier_block profile_exceptions_nb = {
-	.notifier_call = profile_exceptions_notify,
-	.next = NULL,
-	.priority = NMI_LOCAL_LOW_PRIOR,
-};
-
 static void nmi_cpu_restore_registers(struct op_msrs *msrs)
 {
 	struct op_msr *counters = msrs->counters;
@@ -508,7 +491,8 @@ static int nmi_setup(void)
 	ctr_running = 0;
 	/* make variables visible to the nmi handler: */
 	smp_mb();
-	err = register_die_notifier(&profile_exceptions_nb);
+	err = register_nmi_handler(NMI_LOCAL, profile_exceptions_notify,
+					0, "oprofile");
 	if (err)
 		goto fail;
 
@@ -538,7 +522,7 @@ static void nmi_shutdown(void)
 	put_online_cpus();
 	/* make variables visible to the nmi handler: */
 	smp_mb();
-	unregister_die_notifier(&profile_exceptions_nb);
+	unregister_nmi_handler(NMI_LOCAL, "oprofile");
 	msrs = &get_cpu_var(cpu_msrs);
 	model->shutdown(msrs);
 	free_msrs();
arch/x86/oprofile/nmi_timer_int.c +6 -22
index 720bf5a..7f8052c 100644
@@ -18,32 +18,16 @@
 #include <asm/apic.h>
 #include <asm/ptrace.h>
 
-static int profile_timer_exceptions_notify(struct notifier_block *self,
-					   unsigned long val, void *data)
+static int profile_timer_exceptions_notify(unsigned int val, struct pt_regs *regs)
 {
-	struct die_args *args = (struct die_args *)data;
-	int ret = NOTIFY_DONE;
-
-	switch (val) {
-	case DIE_NMI:
-		oprofile_add_sample(args->regs, 0);
-		ret = NOTIFY_STOP;
-		break;
-	default:
-		break;
-	}
-	return ret;
+	oprofile_add_sample(regs, 0);
+	return NMI_HANDLED;
 }
 
-static struct notifier_block profile_timer_exceptions_nb = {
-	.notifier_call = profile_timer_exceptions_notify,
-	.next = NULL,
-	.priority = NMI_LOW_PRIOR,
-};
-
 static int timer_start(void)
 {
-	if (register_die_notifier(&profile_timer_exceptions_nb))
+	if (register_nmi_handler(NMI_LOCAL, profile_timer_exceptions_notify,
+					0, "oprofile-timer"))
 		return 1;
 	return 0;
 }
@@ -51,7 +35,7 @@ static int timer_start(void)
 
 static void timer_stop(void)
 {
-	unregister_die_notifier(&profile_timer_exceptions_nb);
+	unregister_nmi_handler(NMI_LOCAL, "oprofile-timer");
 	synchronize_sched();  /* Allow already-started NMIs to complete. */
 }
 
drivers/acpi/apei/ghes.c +8 -14
index 0784f99..b8e08cb 100644
@@ -50,6 +50,7 @@
 #include <acpi/hed.h>
 #include <asm/mce.h>
 #include <asm/tlbflush.h>
+#include <asm/nmi.h>
 
 #include "apei-internal.h"
 
@@ -749,15 +750,11 @@ static void ghes_proc_in_irq(struct irq_work *irq_work)
 	}
 }
 
-static int ghes_notify_nmi(struct notifier_block *this,
-				  unsigned long cmd, void *data)
+static int ghes_notify_nmi(unsigned int cmd, struct pt_regs *regs)
 {
 	struct ghes *ghes, *ghes_global = NULL;
 	int sev, sev_global = -1;
-	int ret = NOTIFY_DONE;
-
-	if (cmd != DIE_NMI)
-		return ret;
+	int ret = NMI_DONE;
 
 	raw_spin_lock(&ghes_nmi_lock);
 	list_for_each_entry_rcu(ghes, &ghes_nmi, list) {
@@ -770,10 +767,10 @@ static int ghes_notify_nmi(struct notifier_block *this,
 			sev_global = sev;
 			ghes_global = ghes;
 		}
-		ret = NOTIFY_STOP;
+		ret = NMI_HANDLED;
 	}
 
-	if (ret == NOTIFY_DONE)
+	if (ret == NMI_DONE)
 		goto out;
 
 	if (sev_global >= GHES_SEV_PANIC) {
@@ -825,10 +822,6 @@ static struct notifier_block ghes_notifier_sci = {
 	.notifier_call = ghes_notify_sci,
 };
 
-static struct notifier_block ghes_notifier_nmi = {
-	.notifier_call = ghes_notify_nmi,
-};
-
 static unsigned long ghes_esource_prealloc_size(
 	const struct acpi_hest_generic *generic)
 {
@@ -918,7 +911,8 @@ static int __devinit ghes_probe(struct platform_device *ghes_dev)
 		ghes_estatus_pool_expand(len);
 		mutex_lock(&ghes_list_mutex);
 		if (list_empty(&ghes_nmi))
-			register_die_notifier(&ghes_notifier_nmi);
+			register_nmi_handler(NMI_LOCAL, ghes_notify_nmi, 0,
+						"ghes");
 		list_add_rcu(&ghes->list, &ghes_nmi);
 		mutex_unlock(&ghes_list_mutex);
 		break;
@@ -964,7 +958,7 @@ static int __devexit ghes_remove(struct platform_device *ghes_dev)
 		mutex_lock(&ghes_list_mutex);
 		list_del_rcu(&ghes->list);
 		if (list_empty(&ghes_nmi))
-			unregister_die_notifier(&ghes_notifier_nmi);
+			unregister_nmi_handler(NMI_LOCAL, "ghes");
 		mutex_unlock(&ghes_list_mutex);
 		/*
 		 * To synchronize with NMI handler, ghes can only be
drivers/char/ipmi/ipmi_watchdog.c +10 -22
index 3302586..3dcab56 100644
@@ -1077,17 +1077,8 @@ static void ipmi_unregister_watchdog(int ipmi_intf)
 
 #ifdef HAVE_DIE_NMI
 static int
-ipmi_nmi(struct notifier_block *self, unsigned long val, void *data)
+ipmi_nmi(unsigned int val, struct pt_regs *regs)
 {
-	struct die_args *args = data;
-
-	if (val != DIE_NMIUNKNOWN)
-		return NOTIFY_OK;
-
-	/* Hack, if it's a memory or I/O error, ignore it. */
-	if (args->err & 0xc0)
-		return NOTIFY_OK;
-
 	/*
 	 * If we get here, it's an NMI that's not a memory or I/O
 	 * error.  We can't truly tell if it's from IPMI or not
@@ -1097,15 +1088,15 @@ ipmi_nmi(struct notifier_block *self, unsigned long val, void *data)
 
 	if (testing_nmi) {
 		testing_nmi = 2;
-		return NOTIFY_STOP;
+		return NMI_HANDLED;
 	}
 
 	/* If we are not expecting a timeout, ignore it. */
 	if (ipmi_watchdog_state == WDOG_TIMEOUT_NONE)
-		return NOTIFY_OK;
+		return NMI_DONE;
 
 	if (preaction_val != WDOG_PRETIMEOUT_NMI)
-		return NOTIFY_OK;
+		return NMI_DONE;
 
 	/*
 	 * If no one else handled the NMI, we assume it was the IPMI
@@ -1120,12 +1111,8 @@ ipmi_nmi(struct notifier_block *self, unsigned long val, void *data)
 			panic(PFX "pre-timeout");
 	}
 
-	return NOTIFY_STOP;
+	return NMI_HANDLED;
 }
-
-static struct notifier_block ipmi_nmi_handler = {
-	.notifier_call = ipmi_nmi
-};
 #endif
 
 static int wdog_reboot_handler(struct notifier_block *this,
@@ -1290,7 +1277,8 @@ static void check_parms(void)
 		}
 	}
 	if (do_nmi && !nmi_handler_registered) {
-		rv = register_die_notifier(&ipmi_nmi_handler);
+		rv = register_nmi_handler(NMI_UNKNOWN, ipmi_nmi, 0,
+						"ipmi");
 		if (rv) {
 			printk(KERN_WARNING PFX
 			       "Can't register nmi handler\n");
@@ -1298,7 +1286,7 @@ static void check_parms(void)
 		} else
 			nmi_handler_registered = 1;
 	} else if (!do_nmi && nmi_handler_registered) {
-		unregister_die_notifier(&ipmi_nmi_handler);
+		unregister_nmi_handler(NMI_UNKNOWN, "ipmi");
 		nmi_handler_registered = 0;
 	}
 #endif
@@ -1336,7 +1324,7 @@ static int __init ipmi_wdog_init(void)
 	if (rv) {
 #ifdef HAVE_DIE_NMI
 		if (nmi_handler_registered)
-			unregister_die_notifier(&ipmi_nmi_handler);
+			unregister_nmi_handler(NMI_UNKNOWN, "ipmi");
 #endif
 		atomic_notifier_chain_unregister(&panic_notifier_list,
 						 &wdog_panic_notifier);
@@ -1357,7 +1345,7 @@ static void __exit ipmi_wdog_exit(void)
 
 #ifdef HAVE_DIE_NMI
 	if (nmi_handler_registered)
-		unregister_die_notifier(&ipmi_nmi_handler);
+		unregister_nmi_handler(NMI_UNKNOWN, "ipmi");
 #endif
 
 	atomic_notifier_chain_unregister(&panic_notifier_list,
drivers/watchdog/hpwdt.c +6 -17
index 410fba4..e0f6202 100644
@@ -477,15 +477,12 @@ static int hpwdt_time_left(void)
 /*
  *	NMI Handler
  */
-static int hpwdt_pretimeout(struct notifier_block *nb, unsigned long ulReason,
+static int hpwdt_pretimeout(unsigned int ulReason, struct pt_regs *regs,
 				void *data)
 {
 	unsigned long rom_pl;
 	static int die_nmi_called;
 
-	if (ulReason != DIE_NMIUNKNOWN)
-		goto out;
-
 	if (!hpwdt_nmi_decoding)
 		goto out;
 
@@ -507,7 +504,7 @@ static int hpwdt_pretimeout(struct notifier_block *nb, unsigned long ulReason,
 		"Management Log for details.\n");
 
 out:
-	return NOTIFY_OK;
+	return NMI_DONE;
 }
 #endif /* CONFIG_HPWDT_NMI_DECODING */
 
@@ -647,13 +644,6 @@ static struct miscdevice hpwdt_miscdev = {
 	.fops = &hpwdt_fops,
 };
 
-#ifdef CONFIG_HPWDT_NMI_DECODING
-static struct notifier_block die_notifier = {
-	.notifier_call = hpwdt_pretimeout,
-	.priority = 0,
-};
-#endif /* CONFIG_HPWDT_NMI_DECODING */
-
 /*
  *	Init & Exit
  */
@@ -739,10 +729,9 @@ static int __devinit hpwdt_init_nmi_decoding(struct pci_dev *dev)
 	 * die notify list to handle a critical NMI. The default is to
 	 * be last so other users of the NMI signal can function.
 	 */
-	if (priority)
-		die_notifier.priority = 0x7FFFFFFF;
-
-	retval = register_die_notifier(&die_notifier);
+	retval = register_nmi_handler(NMI_UNKNOWN, hpwdt_pretimeout,
+					(priority) ? NMI_HANDLE_FIRST : 0,
+					"hpwdt");
 	if (retval != 0) {
 		dev_warn(&dev->dev,
 			"Unable to register a die notifier (err=%d).\n",
@@ -762,7 +751,7 @@ static int __devinit hpwdt_init_nmi_decoding(struct pci_dev *dev)
 
 static void hpwdt_exit_nmi_decoding(void)
 {
-	unregister_die_notifier(&die_notifier);
+	unregister_nmi_handler(NMI_UNKNOWN, "hpwdt");
 	if (cru_rom_addr)
 		iounmap(cru_rom_addr);
 }
Signature · 1 line
1.7.6
#6 lore raw

Re: [V5][PATCH 3/6] x86, nmi: wire up NMI handlers to new routines

Quoted 5 lines · Don Zickus

Just convert all the files that have an nmi handler to the new routines. Most of it is straight forward conversion. A couple of places needed some tweaking like kgdb which separates the debug notifier from the nmi handler and mce removes a call to notify_die (as I couldn't figure out why it was there).

It is used to call a debugger on a machine check, according to following thread:

https://lkml.org/lkml/2010/5/27/114

So maybe we can turn that into a kgdb direct call?

Best Regards,
Huang Ying

#7 lore raw

Re: [V5][PATCH 3/6] x86, nmi: wire up NMI handlers to new routines

+ Tony.

Quoted 11 lines · Huang Ying
Quoted 5 lines · Don Zickus

Just convert all the files that have an nmi handler to the new routines. Most of it is straight forward conversion. A couple of places needed some tweaking like kgdb which separates the debug notifier from the nmi handler and mce removes a call to notify_die (as I couldn't figure out why it was there).

It is used to call a debugger on a machine check, according to following thread:

https://lkml.org/lkml/2010/5/27/114

Thanks for digging that out - I couldn't find anywhere in the git logs why was this added in the first place.

Quoted 1 line

So maybe we can turn that into a kgdb direct call?

After reading the thread, the semi-legitimate usage of using it as a jump into the debugger just because some hardware reports certain conditions through an MCE sounds pretty hacky to me.

Besides, if the driver developer needs that, he can add the code for the duration of her/his development cycle as aid, and remove it in the end.

This early-exit deal is especially inacceptable if you get an uncorrectable error and some notifier call in the chain consumes it and we never get to report it or decode it, or do recovery action. And thus the box merrily continues on although a corruption just happened and we didn't even get a chance to panic.

So I really really want to remove it, actually.

Signature · 8 lines
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551
author
#8 lore raw

Re: [V5][PATCH 3/6] x86, nmi: wire up NMI handlers to new routines

Quoted 27 lines · Borislav Petkov

+ Tony.

Quoted 11 lines · Huang Ying
Quoted 5 lines · Don Zickus

Just convert all the files that have an nmi handler to the new routines. Most of it is straight forward conversion. A couple of places needed some tweaking like kgdb which separates the debug notifier from the nmi handler and mce removes a call to notify_die (as I couldn't figure out why it was there).

It is used to call a debugger on a machine check, according to following thread:

https://lkml.org/lkml/2010/5/27/114

Thanks for digging that out - I couldn't find anywhere in the git logs why was this added in the first place.

Quoted 1 line

So maybe we can turn that into a kgdb direct call?

After reading the thread, the semi-legitimate usage of using it as a jump into the debugger just because some hardware reports certain conditions through an MCE sounds pretty hacky to me.

Besides, if the driver developer needs that, he can add the code for the duration of her/his development cycle as aid, and remove it in the end.

Quoted 7 lines

This early-exit deal is especially inacceptable if you get an uncorrectable error and some notifier call in the chain consumes it and we never get to report it or decode it, or do recovery action. And thus the box merrily continues on although a corruption just happened and we didn't even get a chance to panic.

So I really really want to remove it, actually.

Cool. Thanks Ying and Boris for settling that. I was scratching my head trying to understand why that was there. It keeps the code simpler to now. :-)

Cheers,
Don

author
#9 lore raw

PATCH v5 4/6 x86, nmi: add in logic to handle multiple events and unknown NMIs

In reply to Don Zickus

Previous patches allow the NMI subsystem to process multipe NMI events in one NMI. As previously discussed this can cause issues when an event triggered another NMI but is processed in the current NMI. This causes the next NMI to go unprocessed and become an 'unknown' NMI.

To handle this, we first have to flag whether or not the NMI handler handled more than one event or not. If it did, then there exists a chance that the next NMI might be already processed. Once the NMI is flagged as a candidate to be swallowed, we next look for a back-to-back NMI condition.

This is determined by looking at the %rip from pt_regs. If it is the same as the previous NMI, it is assumed the cpu did not have a chance to jump back into a non-NMI context and execute code and instead handled another NMI.

If both of those conditions are true then we will swallow any unknown NMI.

There still exists a chance that we accidentally swallow a real unknown NMI, but for now things seem better.

An optimization has also been added to the nmi notifier rountine. Because x86 can latch up to one NMI while currently processing an NMI, we don't have to worry about executing _all_ the handlers in a standalone NMI. The idea is if multiple NMIs come in, the second NMI will represent them. For those back-to-back NMI cases, we have the potentail to drop NMIs. Therefore only execute all the handlers in the second half of a detected back-to-back NMI.

V2:
- forgot to add the 'read' code for swallow_nmi (went into next patch)

V3:
- redesigned the algorithm to utilize Avi's idea of detecting a back-to-back NMI with %rip.
V4:
- clean up fixes, like adding 'static', rename save_rip to last_nmi_rip

---
 arch/x86/kernel/nmi.c |   80 +++++++++++++++++++++++++++++++++++++++++++++---
 1 files changed, 75 insertions(+), 5 deletions(-)
arch/x86/kernel/nmi.c +75 -5
index acd61e8..04c9339 100644
@@ -66,7 +66,7 @@ __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 
 #define nmi_to_desc(type) (&nmi_desc[type])
 
-static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs)
+static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs, bool b2b)
 {
 	struct nmi_desc *desc = nmi_to_desc(type);
 	struct nmiaction *next_a, *a, **ap = &desc->head;
@@ -87,6 +87,16 @@ static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs)
 
 		handled += a->handler(type, regs);
 
+		/*
+ 		 * Optimization: only loop once if this is not a 
+ 		 * back-to-back NMI.  The idea is nothing is dropped
+ 		 * on the first NMI, only on the second of a back-to-back
+ 		 * NMI.  No need to waste cycles going through all the
+ 		 * handlers.
+ 		 */
+		if (!b2b && handled)
+			break;
+
 		a = next_a;
 	}
 	rcu_read_unlock();
@@ -255,7 +265,13 @@ unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
 {
 	int handled;
 
-	handled = nmi_handle(NMI_UNKNOWN, regs);
+	/*
+	 * Use 'false' as back-to-back NMIs are dealt with one level up.
+	 * Of course this makes having multiple 'unknown' handlers useless
+	 * as only the first one is ever run (unless it can actually determine
+	 * if it caused the NMI)
+	 */
+	handled = nmi_handle(NMI_UNKNOWN, regs, false);
 	if (handled) 
 		return;
 #ifdef CONFIG_MCA
@@ -278,19 +294,49 @@ unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
 	pr_emerg("Dazed and confused, but trying to continue\n");
 }
 
+static DEFINE_PER_CPU(bool, swallow_nmi);
+static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
+
 static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
 {
 	unsigned char reason = 0;
 	int handled;
+	bool b2b = false;
 
 	/*
 	 * CPU-specific NMI must be processed before non-CPU-specific
 	 * NMI, otherwise we may lose it, because the CPU-specific
 	 * NMI can not be detected/processed on other CPUs.
 	 */
-	handled = nmi_handle(NMI_LOCAL, regs);
-	if (handled)
+
+	/*
+	 * Back-to-back NMIs are interesting because they can either
+	 * be two NMI or more than two NMIs (any thing over two is dropped
+	 * due to NMI being edge-triggered).  If this is the second half
+	 * of the back-to-back NMI, assume we dropped things and process
+	 * more handlers.  Otherwise reset the 'swallow' NMI behaviour
+	 */
+	if (regs->ip == __this_cpu_read(last_nmi_rip))
+		b2b = true;
+	else
+		__this_cpu_write(swallow_nmi, false);
+
+	__this_cpu_write(last_nmi_rip, regs->ip);
+
+	handled = nmi_handle(NMI_LOCAL, regs, b2b);
+	if (handled) {
+		/*
+ 		 * There are cases when a NMI handler handles multiple
+ 		 * events in the current NMI.  One of these events may
+ 		 * be queued for in the next NMI.  Because the event is
+ 		 * already handled, the next NMI will result in an unknown
+ 		 * NMI.  Instead lets flag this for a potential NMI to
+ 		 * swallow.
+		 */
+		if (handled > 1)
+			__this_cpu_write(swallow_nmi, true);
 		return;
+	}
 
 	/* Non-CPU-specific NMI: NMI sources can be processed on any CPU */
 	raw_spin_lock(&nmi_reason_lock);
@@ -313,7 +359,31 @@ static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
 	}
 	raw_spin_unlock(&nmi_reason_lock);
 
-	unknown_nmi_error(reason, regs);
+	/*
+	 * Only one NMI can be latched at a time.  To handle
+	 * this we may process multiple nmi handlers at once to
+	 * cover the case where an NMI is dropped.  The downside
+	 * to this approach is we may process an NMI prematurely,
+	 * while its real NMI is sitting latched.  This will cause
+	 * an unknown NMI on the next run of the NMI processing.
+	 * 
+	 * We tried to flag that condition above, by setting the
+	 * swallow_nmi flag when we process more than one event.
+	 * This condition is also only present on the second half
+	 * of a back-to-back NMI, so we flag that condition too.
+	 *
+	 * If both are true, we assume we already processed this
+	 * NMI previously and we swallow it.  Otherwise we reset
+	 * the logic.
+	 *
+	 * I am sure there are scenarios where we accidentally
+	 * swallow a real 'unknown' NMI.  But this is the best
+	 * we can do for now.
+	 */
+	if (b2b && __this_cpu_read(swallow_nmi))
+		;
+	else
+		unknown_nmi_error(reason, regs);
 }
 
 dotraplinkage notrace __kprobes void
Signature · 1 line
1.7.6
#10 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 26 lines · Don Zickus

Previous patches allow the NMI subsystem to process multipe NMI events in one NMI. As previously discussed this can cause issues when an event triggered another NMI but is processed in the current NMI. This causes the next NMI to go unprocessed and become an 'unknown' NMI.

To handle this, we first have to flag whether or not the NMI handler handled more than one event or not. If it did, then there exists a chance that the next NMI might be already processed. Once the NMI is flagged as a candidate to be swallowed, we next look for a back-to-back NMI condition.

This is determined by looking at the %rip from pt_regs. If it is the same as the previous NMI, it is assumed the cpu did not have a chance to jump back into a non-NMI context and execute code and instead handled another NMI.

If both of those conditions are true then we will swallow any unknown NMI.

There still exists a chance that we accidentally swallow a real unknown NMI, but for now things seem better.

An optimization has also been added to the nmi notifier rountine. Because x86 can latch up to one NMI while currently processing an NMI, we don't have to worry about executing _all_ the handlers in a standalone NMI. The idea is if multiple NMIs come in, the second NMI will represent them. For those back-to-back NMI cases, we have the potentail to drop NMIs. Therefore only execute all the handlers in the second half of a detected back-to-back NMI.

Quoted 10 lines

V2:
- forgot to add the 'read' code for swallow_nmi (went into next patch)

V3:
- redesigned the algorithm to utilize Avi's idea of detecting a back-to-back NMI with %rip.
V4:
- clean up fixes, like adding 'static', rename save_rip to last_nmi_rip

Missing a zeroing of last_nmi_rip in the idle path. Otherwise, as Andi points out, and idle machine will always see NMIs coming in from the hlt/mwait address and detect them as back-to-back NMIs.

Signature · 1 line
error compiling committee.c: too many arguments to function
author
#11 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 8 lines · Avi Kivity
Quoted 4 lines

V4:
- clean up fixes, like adding 'static', rename save_rip to last_nmi_rip

Missing a zeroing of last_nmi_rip in the idle path. Otherwise, as Andi points out, and idle machine will always see NMIs coming in from the hlt/mwait address and detect them as back-to-back NMIs.

You mean something like this?

From: Don Zickus <[email protected]>
Date: Fri, 19 Aug 2011 15:51:44 -0400
Subject: [PATCH] x86, nmi: add in logic to handle multiple events and
unknown NMIs

Previous patches allow the NMI subsystem to process multipe NMI events in one NMI. As previously discussed this can cause issues when an event triggered another NMI but is processed in the current NMI. This causes the next NMI to go unprocessed and become an 'unknown' NMI.

To handle this, we first have to flag whether or not the NMI handler handled more than one event or not. If it did, then there exists a chance that the next NMI might be already processed. Once the NMI is flagged as a candidate to be swallowed, we next look for a back-to-back NMI condition.

This is determined by looking at the %rip from pt_regs. If it is the same as the previous NMI, it is assumed the cpu did not have a chance to jump back into a non-NMI context and execute code and instead handled another NMI.

If both of those conditions are true then we will swallow any unknown NMI.

There still exists a chance that we accidentally swallow a real unknown NMI, but for now things seem better.

An optimization has also been added to the nmi notifier rountine. Because x86 can latch up to one NMI while currently processing an NMI, we don't have to worry about executing _all_ the handlers in a standalone NMI. The idea is if multiple NMIs come in, the second NMI will represent them. For those back-to-back NMI cases, we have the potentail to drop NMIs. Therefore only execute all the handlers in the second half of a detected back-to-back NMI.

V2:
- forgot to add the 'read' code for swallow_nmi (went into next patch)

V3:
- redesigned the algorithm to utilize Avi's idea of detecting a back-to-back NMI with %rip.
V4:
- clean up fixes, like adding 'static', rename save_rip to last_nmi_rip

V5:
- wire up the idle path to reset the back-to-back NMI logic

---
 arch/x86/include/asm/nmi.h   |    1 +
 arch/x86/kernel/nmi.c        |   86 +++++++++++++++++++++++++++++++++++++++--
 arch/x86/kernel/process_32.c |    2 +
 arch/x86/kernel/process_64.c |    2 +
 4 files changed, 86 insertions(+), 5 deletions(-)
arch/x86/include/asm/nmi.h +1 -0
index fc74547..0930d4a 100644
@@ -43,5 +43,6 @@ void unregister_nmi_handler(unsigned int, const char *);
 
 void stop_nmi(void);
 void restart_nmi(void);
+void local_touch_nmi(void);
 
 #endif /* _ASM_X86_NMI_H */
arch/x86/kernel/nmi.c +81 -5

130 lines not on this page · raw message

arch/x86/kernel/process_32.c +2 -0
index 7a3b651..46ff054 100644
@@ -57,6 +57,7 @@
 #include <asm/idle.h>
 #include <asm/syscalls.h>
 #include <asm/debugreg.h>
+#include <asm/nmi.h>
 
 asmlinkage void ret_from_fork(void) __asm__("ret_from_fork");
 
@@ -107,6 +108,7 @@ void cpu_idle(void)
 			if (cpu_is_offline(cpu))
 				play_dead();
 
+			local_touch_nmi();
 			local_irq_disable();
 			/* Don't trace irqs off for idle */
 			stop_critical_timings();
arch/x86/kernel/process_64.c +2 -0
index f693e44..3bd7e6e 100644
@@ -51,6 +51,7 @@
 #include <asm/idle.h>
 #include <asm/syscalls.h>
 #include <asm/debugreg.h>
+#include <asm/nmi.h>
 
 asmlinkage extern void ret_from_fork(void);
 
@@ -133,6 +134,7 @@ void cpu_idle(void)
 			 * from here on, until they go to idle.
 			 * Otherwise, idle callbacks can misfire.
 			 */
+			local_touch_nmi();
 			local_irq_disable();
 			enter_idle();
 			/* Don't trace irqs off for idle */
Signature · 1 line
1.7.6
#12 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 6 lines · Don Zickus
Quoted 5 lines · Avi Kivity
Quoted 4 lines

V4:
- clean up fixes, like adding 'static', rename save_rip to last_nmi_rip

Quoted 6 lines
Quoted 3 lines

Missing a zeroing of last_nmi_rip in the idle path. Otherwise, as Andi points out, and idle machine will always see NMIs coming in from the hlt/mwait address and detect them as back-to-back NMIs.

You mean something like this?

Yes.

Signature · 2 lines
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.
#13 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

In reply to Don Zickus

On 09/20/2011 10:43 PM, Don Zickus wrote:
[snip]

Quoted 15 lines
arch/x86/kernel/nmi.c +10 -1
@@ -313,7 +359,31 @@ static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
 	}
 	raw_spin_unlock(&nmi_reason_lock);
 
-	unknown_nmi_error(reason, regs);
+	/*
+	 * Only one NMI can be latched at a time.  To handle
+	 * this we may process multiple nmi handlers at once to
+	 * cover the case where an NMI is dropped.  The downside
+	 * to this approach is we may process an NMI prematurely,
+	 * while its real NMI is sitting latched.  This will cause
+	 * an unknown NMI on the next run of the NMI processing.
+	 * 
+	 * We tried to flag that condition above, by setting the
+	 * swallow_nmi flag when we process more than one event.

arch/x86/kernel/nmi.c static notrace __kprobes void default_do_nmi(struct pt_regs *regs) · line 381

Quoted 10 lines
arch/x86/kernel/nmi.c +10 -0
+	 * This condition is also only present on the second half
+	 * of a back-to-back NMI, so we flag that condition too.
+	 *
+	 * If both are true, we assume we already processed this
+	 * NMI previously and we swallow it.  Otherwise we reset
+	 * the logic.
+	 *
+	 * I am sure there are scenarios where we accidentally
+	 * swallow a real 'unknown' NMI.  But this is the best
+	 * we can do for now.

Why not describe a scenario where we swallow a real 'unknown' NMI? So that someone working on the code in the future will know the challenge?

Best Regards,
Huang Ying

Quoted 8 lines
arch/x86/kernel/nmi.c +5 -0
+	 */
+	if (b2b && __this_cpu_read(swallow_nmi))
+		;
+	else
+		unknown_nmi_error(reason, regs);
 }
 
 dotraplinkage notrace __kprobes void
author
#14 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 22 lines · Huang Ying

On 09/20/2011 10:43 PM, Don Zickus wrote:
[snip]

Quoted 20 lines
arch/x86/kernel/nmi.c +15 -1
@@ -313,7 +359,31 @@ static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
 	}
 	raw_spin_unlock(&nmi_reason_lock);
 
-	unknown_nmi_error(reason, regs);
+	/*
+	 * Only one NMI can be latched at a time.  To handle
+	 * this we may process multiple nmi handlers at once to
+	 * cover the case where an NMI is dropped.  The downside
+	 * to this approach is we may process an NMI prematurely,
+	 * while its real NMI is sitting latched.  This will cause
+	 * an unknown NMI on the next run of the NMI processing.
+	 * 
+	 * We tried to flag that condition above, by setting the
+	 * swallow_nmi flag when we process more than one event.
+	 * This condition is also only present on the second half
+	 * of a back-to-back NMI, so we flag that condition too.
+	 *
+	 * If both are true, we assume we already processed this
+	 * NMI previously and we swallow it.  Otherwise we reset
Quoted 8 lines

arch/x86/kernel/nmi.c static notrace __kprobes void default_do_nmi(struct pt_regs *regs) · line 381

Quoted 5 lines
arch/x86/kernel/nmi.c +5 -0
+	 * the logic.
+	 *
+	 * I am sure there are scenarios where we accidentally
+	 * swallow a real 'unknown' NMI.  But this is the best
+	 * we can do for now.

Why not describe a scenario where we swallow a real 'unknown' NMI? So that someone working on the code in the future will know the challenge?

I can add a couple of lines of comment for that.

Thanks,
Don

#15 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

In reply to Don Zickus

arch/x86/kernel/nmi.c static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs) · line 98

Quoted 13 lines · Don Zickus
arch/x86/kernel/nmi.c +9 -0
@@ -87,6 +87,16 @@ static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs)
 
 		handled += a->handler(type, regs);
 
+		/*
+ 		 * Optimization: only loop once if this is not a 
+ 		 * back-to-back NMI.  The idea is nothing is dropped
+ 		 * on the first NMI, only on the second of a back-to-back
+ 		 * NMI.  No need to waste cycles going through all the
+ 		 * handlers.
+ 		 */
+		if (!b2b && handled)
+			break;

In rare cases we will lose nmis here.

We see a back-to-back nmi in the case if a 2 nmi source triggers *after* the nmi handler is entered. Depending on the internal cpu timing influenced by microcode and SMM code execution, the nmi may not entered immediately. So all sources that trigger *before* the nmi handler is entered raise only one nmi with no subsequent nmi.

However, as these cases should be very rare, I think we can live with it in favor of the optimization to jump out the handler chain and save lot of cpu cycles esp. in the case of heavy PMI load.

-Robert

Signature · 2 lines
Advanced Micro Devices, Inc.
Operating System Research Center
author
#16 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 17 lines · Robert Richter

arch/x86/kernel/nmi.c static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs) · line 98

Quoted 13 lines · Don Zickus
arch/x86/kernel/nmi.c +9 -0
@@ -87,6 +87,16 @@ static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs)
 
 		handled += a->handler(type, regs);
 
+		/*
+ 		 * Optimization: only loop once if this is not a 
+ 		 * back-to-back NMI.  The idea is nothing is dropped
+ 		 * on the first NMI, only on the second of a back-to-back
+ 		 * NMI.  No need to waste cycles going through all the
+ 		 * handlers.
+ 		 */
+		if (!b2b && handled)
+			break;

In rare cases we will lose nmis here.

Quoted 5 lines

We see a back-to-back nmi in the case if a 2 nmi source triggers *after* the nmi handler is entered. Depending on the internal cpu timing influenced by microcode and SMM code execution, the nmi may not entered immediately. So all sources that trigger *before* the nmi handler is entered raise only one nmi with no subsequent nmi.

Right, but that can only happen with the second NMI in the back-to-back NMI case. Here the optimization is only for the first NMI, with the assumption that you will always have a second NMI if multiple sources trigger, so you can process those in the second iteration (assuming we correctly detect the back-to-back NMI condition). Then when the second NMI comes in, we have no idea how many we dropped to get here so we process all the handlers based on the assumption we might not have another NMI behind us to make up for the dropped NMIs.

Unless I misunderstood your point above?

Quoted 4 lines

However, as these cases should be very rare, I think we can live with it in favor of the optimization to jump out the handler chain and save lot of cpu cycles esp. in the case of heavy PMI load.

I think it is covered as described above?

Cheers,
Don

#17 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 33 lines · Don Zickus
Quoted 17 lines · Robert Richter

arch/x86/kernel/nmi.c static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs) · line 98

Quoted 13 lines · Don Zickus
arch/x86/kernel/nmi.c +9 -0
@@ -87,6 +87,16 @@ static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs)
 
 		handled += a->handler(type, regs);
 
+		/*
+ 		 * Optimization: only loop once if this is not a 
+ 		 * back-to-back NMI.  The idea is nothing is dropped
+ 		 * on the first NMI, only on the second of a back-to-back
+ 		 * NMI.  No need to waste cycles going through all the
+ 		 * handlers.
+ 		 */
+		if (!b2b && handled)
+			break;

In rare cases we will lose nmis here.

Quoted 5 lines

We see a back-to-back nmi in the case if a 2 nmi source triggers *after* the nmi handler is entered. Depending on the internal cpu timing influenced by microcode and SMM code execution, the nmi may not entered immediately. So all sources that trigger *before* the nmi handler is entered raise only one nmi with no subsequent nmi.

Right, but that can only happen with the second NMI in the back-to-back NMI case. Here the optimization is only for the first NMI, with the assumption that you will always have a second NMI if multiple sources trigger, so you can process those in the second iteration (assuming we correctly detect the back-to-back NMI condition). Then when the second NMI comes in, we have no idea how many we dropped to get here so we process all the handlers based on the assumption we might not have another NMI behind us to make up for the dropped NMIs.

Quoted 1 line

Unless I misunderstood your point above?

No, my point was that a second NMI might not be latched even if there are two nmi sources pending.

Your logic is correct but assumes you will always receive a second nmi. This is not always the case depending on the cpu's internal timing. Usually there is the following sequence for back-to-back nmis:

1. HW triggers first NMI, an NMI is pending.
2. NMI handler is called, no NMI pending anymore.
3. HW triggers a second NMI, an NMI is pending.
4. Return from NMI handler.
5. NMI handler is called again to serve the 2nd, no NMI pending anymore.
6. Return from NMI handler.

The above is what your algorithm covers.

But in rare cases there is the following:

1. The cpu executes some microcode or SMM code.
2. HW triggers the first NMI, an NMI is pending.
3. HW triggers a second NMI, the NMI is still pending.
4. The cpu finished microcode or SMM code.
5. NMI handler is called, no NMI pending anymore.
6. Return from NMI handler.

In this case the handler is called only once and the second nmi remains unhandled with you implementation.

I don't see a way how this could be catched without serving all handlers the first time. But as said, in favor of the optimization I think we can live with losing some NMIs.

-Robert

Quoted 3 lines
Quoted 3 lines

However, as these cases should be very rare, I think we can live with it in favor of the optimization to jump out the handler chain and save lot of cpu cycles esp. in the case of heavy PMI load.

Signature · 2 lines
Advanced Micro Devices, Inc.
Operating System Research Center
maintainer
#18 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 6 lines · Robert Richter

1. The cpu executes some microcode or SMM code.
2. HW triggers the first NMI, an NMI is pending.
3. HW triggers a second NMI, the NMI is still pending.
4. The cpu finished microcode or SMM code.
5. NMI handler is called, no NMI pending anymore.
6. Return from NMI handler.

Even without SMM, all you need is two different NMI users to trigger while an NMI is in flight.

Wouldn't be entirely impossible to trigger if you have non-fatal-MCE/hardware-NMI-switch/PMI all active.

#19 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 8 lines · Peter Zijlstra
Quoted 6 lines · Robert Richter

1. The cpu executes some microcode or SMM code.
2. HW triggers the first NMI, an NMI is pending.
3. HW triggers a second NMI, the NMI is still pending.
4. The cpu finished microcode or SMM code.
5. NMI handler is called, no NMI pending anymore.
6. Return from NMI handler.

Quoted 5 lines

Even without SMM, all you need is two different NMI users to trigger while an NMI is in flight.

Wouldn't be entirely impossible to trigger if you have non-fatal-MCE/hardware-NMI-switch/PMI all active.

No, I am trying to explain this...

If hw triggers more than one nmi *before* the nmi handler is entered (no matter what the source is), then there will be no back-to-back nmi. NMIs are edge triggered and there is only one pending signal (AMD cpus). The signal is cleared when entering the handler. If the cpu executes microcode or smm code and delays execution of the nmi handler, then there is no second call of the handler even if there are 2 sources for it.

-Robert

Signature · 2 lines
Advanced Micro Devices, Inc.
Operating System Research Center
maintainer
#20 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 1 line · Robert Richter

No, I am trying to explain this...

D'uh, I got the b2b case inverted.. you're right.

author
#21 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

In reply to Robert Richter
Quoted 61 lines · Robert Richter
Quoted 33 lines · Don Zickus
Quoted 17 lines · Robert Richter

arch/x86/kernel/nmi.c static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs) · line 98

Quoted 13 lines · Don Zickus
arch/x86/kernel/nmi.c +9 -0
@@ -87,6 +87,16 @@ static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs)
 
 		handled += a->handler(type, regs);
 
+		/*
+ 		 * Optimization: only loop once if this is not a 
+ 		 * back-to-back NMI.  The idea is nothing is dropped
+ 		 * on the first NMI, only on the second of a back-to-back
+ 		 * NMI.  No need to waste cycles going through all the
+ 		 * handlers.
+ 		 */
+		if (!b2b && handled)
+			break;

In rare cases we will lose nmis here.

Quoted 5 lines

We see a back-to-back nmi in the case if a 2 nmi source triggers *after* the nmi handler is entered. Depending on the internal cpu timing influenced by microcode and SMM code execution, the nmi may not entered immediately. So all sources that trigger *before* the nmi handler is entered raise only one nmi with no subsequent nmi.

Right, but that can only happen with the second NMI in the back-to-back NMI case. Here the optimization is only for the first NMI, with the assumption that you will always have a second NMI if multiple sources trigger, so you can process those in the second iteration (assuming we correctly detect the back-to-back NMI condition). Then when the second NMI comes in, we have no idea how many we dropped to get here so we process all the handlers based on the assumption we might not have another NMI behind us to make up for the dropped NMIs.

Quoted 1 line

Unless I misunderstood your point above?

No, my point was that a second NMI might not be latched even if there are two nmi sources pending.

Your logic is correct but assumes you will always receive a second nmi. This is not always the case depending on the cpu's internal timing. Usually there is the following sequence for back-to-back nmis:

1. HW triggers first NMI, an NMI is pending.
2. NMI handler is called, no NMI pending anymore.
3. HW triggers a second NMI, an NMI is pending.
4. Return from NMI handler.
5. NMI handler is called again to serve the 2nd, no NMI pending anymore.
6. Return from NMI handler.

The above is what your algorithm covers.

But in rare cases there is the following:

1. The cpu executes some microcode or SMM code.
2. HW triggers the first NMI, an NMI is pending.
3. HW triggers a second NMI, the NMI is still pending.
4. The cpu finished microcode or SMM code.
5. NMI handler is called, no NMI pending anymore.
6. Return from NMI handler.

Quoted 6 lines

In this case the handler is called only once and the second nmi remains unhandled with you implementation.

I don't see a way how this could be catched without serving all handlers the first time. But as said, in favor of the optimization I think we can live with losing some NMIs.

Ah, I get it know. Crap. Well I think Avi was pushing it to make those ticket_spin_locks work in virt land. It seems like we should lean towards removing the optimization. Avi?

Cheers,
Don

#22 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 65 lines · Don Zickus
Quoted 64 lines · Robert Richter
Quoted 33 lines · Don Zickus
Quoted 15 lines · Robert Richter
Quoted 4 lines · Don Zickus

@@ -87,6 +87,16 @@ static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs)

  		handled += a->handler(type, regs);
Quoted 9 lines

+ /*
+ * Optimization: only loop once if this is not a
+ * back-to-back NMI. The idea is nothing is dropped
+ * on the first NMI, only on the second of a back-to-back
+ * NMI. No need to waste cycles going through all the
+ * handlers.
+ */
+ if (!b2b&& handled)
+ break;

Quoted 7 lines

In rare cases we will lose nmis here.

We see a back-to-back nmi in the case if a 2 nmi source triggers *after* the nmi handler is entered. Depending on the internal cpu timing influenced by microcode and SMM code execution, the nmi may not entered immediately. So all sources that trigger *before* the nmi handler is entered raise only one nmi with no subsequent nmi.

Right, but that can only happen with the second NMI in the back-to-back NMI case. Here the optimization is only for the first NMI, with the assumption that you will always have a second NMI if multiple sources trigger, so you can process those in the second iteration (assuming we correctly detect the back-to-back NMI condition). Then when the second NMI comes in, we have no idea how many we dropped to get here so we process all the handlers based on the assumption we might not have another NMI behind us to make up for the dropped NMIs.

Quoted 1 line

Unless I misunderstood your point above?

No, my point was that a second NMI might not be latched even if there are two nmi sources pending.

Your logic is correct but assumes you will always receive a second nmi. This is not always the case depending on the cpu's internal timing. Usually there is the following sequence for back-to-back nmis:

1. HW triggers first NMI, an NMI is pending.
2. NMI handler is called, no NMI pending anymore.
3. HW triggers a second NMI, an NMI is pending.
4. Return from NMI handler.
5. NMI handler is called again to serve the 2nd, no NMI pending anymore.
6. Return from NMI handler.

The above is what your algorithm covers.

But in rare cases there is the following:

1. The cpu executes some microcode or SMM code.
2. HW triggers the first NMI, an NMI is pending.
3. HW triggers a second NMI, the NMI is still pending.
4. The cpu finished microcode or SMM code.
5. NMI handler is called, no NMI pending anymore.
6. Return from NMI handler.

In this case the handler is called only once and the second nmi remains unhandled with you implementation.

Quoted 8 lines
Quoted 3 lines

I don't see a way how this could be catched without serving all handlers the first time. But as said, in favor of the optimization I think we can live with losing some NMIs.

Ah, I get it know. Crap. Well I think Avi was pushing it to make those ticket_spin_locks work in virt land. It seems like we should lean towards removing the optimization. Avi?

Well, in virt land there are no SMIs, and we can guarantee that the queue length is always two. So if these rare cases are okay for upstream, it'll be fine for virt.

Signature · 1 line
error compiling committee.c: too many arguments to function
#23 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 13 lines · Avi Kivity
Quoted 12 lines · Don Zickus
Quoted 11 lines · Robert Richter

On 21.09.11 10:04:32, Don Zickus wrote:

But in rare cases there is the following:

1. The cpu executes some microcode or SMM code.
2. HW triggers the first NMI, an NMI is pending.
3. HW triggers a second NMI, the NMI is still pending.
4. The cpu finished microcode or SMM code.
5. NMI handler is called, no NMI pending anymore.
6. Return from NMI handler.

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

In this case the handler is called only once and the second nmi remains unhandled with you implementation.

I don't see a way how this could be catched without serving all handlers the first time. But as said, in favor of the optimization I think we can live with losing some NMIs.

I have to revise this after thinking more about this. We may not lose an nmi for sources where the nmi handler must always reenable the nmi, e.g. IBS. Losing one nmi means for IBS that sample generation gets stuck.

-Robert

Quoted 14 lines
Quoted 5 lines

Ah, I get it know. Crap. Well I think Avi was pushing it to make those ticket_spin_locks work in virt land. It seems like we should lean towards removing the optimization. Avi?

Well, in virt land there are no SMIs, and we can guarantee that the queue length is always two. So if these rare cases are okay for upstream, it'll be fine for virt.

Signature · 1 line
error compiling committee.c: too many arguments to function
Signature · 2 lines
Advanced Micro Devices, Inc.
Operating System Research Center
#24 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

Quoted 17 lines · Robert Richter
Quoted 16 lines · Avi Kivity
Quoted 15 lines · Don Zickus
Quoted 14 lines · Robert Richter

On 21.09.11 10:04:32, Don Zickus wrote:

But in rare cases there is the following:

1. The cpu executes some microcode or SMM code.
2. HW triggers the first NMI, an NMI is pending.
3. HW triggers a second NMI, the NMI is still pending.
4. The cpu finished microcode or SMM code.
5. NMI handler is called, no NMI pending anymore.
6. Return from NMI handler.

In this case the handler is called only once and the second nmi remains unhandled with you implementation.

Quoted 9 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

I don't see a way how this could be catched without serving all handlers the first time. But as said, in favor of the optimization I think we can live with losing some NMIs.

I have to revise this after thinking more about this. We may not lose an nmi for sources where the nmi handler must always reenable the nmi, e.g. IBS. Losing one nmi means for IBS that sample generation gets stuck.

Well, that pretty much kills the whole idea. This thing has to be reliable.

I'll ask Intel if they can guarantee a length 2 queue on their processors (or maybe Andi you can find this out).

Signature · 1 line
error compiling committee.c: too many arguments to function
author
#25 lore raw

Re: [V5][PATCH 4/6] x86, nmi: add in logic to handle multiple events and unknown NMIs

In reply to Avi Kivity
Quoted 16 lines · Avi Kivity
Quoted 16 lines
Quoted 15 lines

But in rare cases there is the following:

1. The cpu executes some microcode or SMM code.
2. HW triggers the first NMI, an NMI is pending.
3. HW triggers a second NMI, the NMI is still pending.
4. The cpu finished microcode or SMM code.
5. NMI handler is called, no NMI pending anymore.
6. Return from NMI handler.

In this case the handler is called only once and the second nmi remains unhandled with you implementation.

I don't see a way how this could be catched without serving all handlers the first time. But as said, in favor of the optimization I think we can live with losing some NMIs.

Quoted 8 lines
Quoted 4 lines

Ah, I get it know. Crap. Well I think Avi was pushing it to make those ticket_spin_locks work in virt land. It seems like we should lean towards removing the optimization. Avi?

Well, in virt land there are no SMIs, and we can guarantee that the queue length is always two. So if these rare cases are okay for upstream, it'll be fine for virt.

Actually I was trying to remember what the argument for the optimization was again? I believe the virt team needed it right?

Cheers,
Don

author
#26 lore raw

PATCH v5 5/6 x86, nmi: track NMI usage stats

In reply to Don Zickus

Now that the NMI handler are broken into lists, increment the appropriate stats for each list. This allows us to see what is going on when they get printed out in the next patch.

---
 arch/x86/kernel/nmi.c |   20 ++++++++++++++++++--
 1 files changed, 18 insertions(+), 2 deletions(-)
arch/x86/kernel/nmi.c +18 -2
index 04c9339..bb8242e 100644
@@ -48,6 +48,15 @@ static struct nmi_desc nmi_desc[NMI_MAX] =
 
 };
 
+struct nmi_stats {
+	unsigned int normal;
+	unsigned int unknown;
+	unsigned int external;
+	unsigned int swallow;
+};
+
+static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
+
 static int ignore_nmis;
 
 int unknown_nmi_panic;
@@ -272,8 +281,13 @@ unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
 	 * if it caused the NMI)
 	 */
 	handled = nmi_handle(NMI_UNKNOWN, regs, false);
-	if (handled) 
+	if (handled) {
+		__this_cpu_add(nmi_stats.unknown, handled);
 		return;
+	}
+
+	__this_cpu_add(nmi_stats.unknown, 1);
+
 #ifdef CONFIG_MCA
 	/*
 	 * Might actually be able to figure out what the guilty party
@@ -324,6 +338,7 @@ static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
 	__this_cpu_write(last_nmi_rip, regs->ip);
 
 	handled = nmi_handle(NMI_LOCAL, regs, b2b);
+	__this_cpu_add(nmi_stats.normal, handled);
 	if (handled) {
 		/*
  		 * There are cases when a NMI handler handles multiple
@@ -354,6 +369,7 @@ static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
 		 */
 		reassert_nmi();
 #endif
+		__this_cpu_add(nmi_stats.external, 1);
 		raw_spin_unlock(&nmi_reason_lock);
 		return;
 	}
@@ -381,7 +397,7 @@ static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
 	 * we can do for now.
 	 */
 	if (b2b && __this_cpu_read(swallow_nmi))
-		;
+		__this_cpu_add(nmi_stats.swallow, 1);
 	else
 		unknown_nmi_error(reason, regs);
 }
Signature · 1 line
1.7.6
author
#27 lore raw

PATCH v5 6/6 x86, nmi: print out NMI stats in /proc/interrupts

In reply to Don Zickus

This is a cheap hack to add the stats to the middle of /proc/interrupts. It is more of a conversation starter than anything as I am not sure the right letters and place to put this stuff.

The benefit of these stats is a better breakdown of which list the NMIs get handled in either a normal handler, unknown, or external. It also list the number of unknown NMIs swallowed to help check for false positives or not. Another benefit is the ability to actually see which NMI handlers are currently registered in the system.

The output of 'cat /proc/interrupts/ will look like this:

<snip>

 58:        275          0        864          0   PCI-MSI-edge      eth0
NMI:       4161       4155        158       4194   Non-maskable interrupts
SWA:          0          0          0          0   Unknown NMIs swallowed
  0:       4161       4155        158       4194   NMI  PMI, arch_bt
UNK:          0          0          0          0   NMI
EXT:          0          0          0          0   NMI
LOC:      12653      13304      13974      12926   Local timer interrupts
SPU:          0          0          0          0   Spurious interrupts
PMI:          6          6          5          6   Performance monitoring interrupts
IWI:          0          0          0          0   IRQ work interrupts
RES:       1839       1897       1821       1854   Rescheduling interrupts
CAL:        524       2714        392        331   Function call interrupts
TLB:        217        146        593        576   TLB shootdowns
TRM:          0          0          0          0   Thermal event interrupts
THR:          0          0          0          0   Threshold APIC interrupts
MCE:          0          0          0          0   Machine check exceptions
MCP:          1          1          1          1   Machine check polls
ERR:          0
MIS:          0
---
 arch/x86/include/asm/nmi.h |    2 +
 arch/x86/kernel/irq.c      |    2 +
 arch/x86/kernel/nmi.c      |   47 ++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 51 insertions(+), 0 deletions(-)
arch/x86/include/asm/nmi.h +2 -0
index fc74547..a4f1945 100644
@@ -24,6 +24,8 @@ void arch_trigger_all_cpu_backtrace(void);
 
 #define NMI_FLAG_FIRST	1
 
+void arch_show_nmi(struct seq_file *p, int prec);
+
 enum {
 	NMI_LOCAL=0,
 	NMI_UNKNOWN,
arch/x86/kernel/irq.c +2 -0
index 6c0802e..44d1cac 100644
@@ -16,6 +16,7 @@
 #include <asm/idle.h>
 #include <asm/mce.h>
 #include <asm/hw_irq.h>
+#include <asm/nmi.h>
 
 atomic_t irq_err_count;
 
@@ -55,6 +56,7 @@ int arch_show_interrupts(struct seq_file *p, int prec)
 	for_each_online_cpu(j)
 		seq_printf(p, "%10u ", irq_stats(j)->__nmi_count);
 	seq_printf(p, "  Non-maskable interrupts\n");
+	arch_show_nmi(p, prec);
 #ifdef CONFIG_X86_LOCAL_APIC
 	seq_printf(p, "%*s: ", prec, "LOC");
 	for_each_online_cpu(j)
arch/x86/kernel/nmi.c +47 -0

50 lines not on this page · raw message

Signature · 1 line
1.7.6