[PATCH 0/5] More NMI IPI enablement work

[PATCH 0/5] More NMI IPI enablement work

[Nicholas Piggin]

Hi,

Here is the NMI IPI enablement for powernv, which requires
supported firmware and hardware:

https://lists.ozlabs.org/pipermail/skiboot/2017-September/009111.html

And some preceeding patches that fix issues I ran into when
testing it.

Thanks,
Nick

Nicholas Piggin (5):
  powerpc/watchdog: do not panic from locked CPU's IPI handler
  powerpc/watchdog: do not backtrace locked CPUs twice if allcpus
    backtrace is enabled
  powerpc/watchdog: do not trigger SMP crash from touch_nmi_watchdog
  powerpc/xmon: avoid tripping SMP hardlockup watchdog
  powerpc/powernv: implement NMI IPI with OPAL_SIGNAL_SYSTEM_RESET

 arch/powerpc/include/asm/opal-api.h            |  1 +
 arch/powerpc/include/asm/opal.h                |  2 ++
 arch/powerpc/kernel/irq.c                      | 43 +++++++++++++++++++---
 arch/powerpc/kernel/watchdog.c                 | 29 ++++++++-------
 arch/powerpc/platforms/powernv/opal-wrappers.S |  1 +
 arch/powerpc/platforms/powernv/powernv.h       |  1 +
 arch/powerpc/platforms/powernv/setup.c         |  3 ++
 arch/powerpc/platforms/powernv/smp.c           | 50 ++++++++++++++++++++++++++
 arch/powerpc/xmon/xmon.c                       | 17 ++++++---
 9 files changed, 127 insertions(+), 20 deletions(-)

-- 
2.13.3

[PATCH 1/5] powerpc/watchdog: do not panic from locked CPU's IPI handler

[Nicholas Piggin]

The SMP watchdog will detect locked CPUs and IPI them to print a
backtrace and registers. If panic on hard lockup is enabled, do
not panic from this handler, because that can cause recursion into
the IPI layer during the panic.

The caller already panics in this case.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/powerpc/kernel/watchdog.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index 2f6eadd9408d..532a1adbe89b 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c
@@ -97,8 +97,7 @@ static void wd_lockup_ipi(struct pt_regs *regs)
 	else
 		dump_stack();
 
-	if (hardlockup_panic)
-		nmi_panic(regs, "Hard LOCKUP");
+	/* Do not panic from here because that can recurse into NMI IPI layer */
 }
 
 static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
-- 
2.13.3

[PATCH 2/5] powerpc/watchdog: do not backtrace locked CPUs twice if allcpus backtrace is enabled

[Nicholas Piggin]

If sysctl_hardlockup_all_cpu_backtrace is enabled, there is no need to
IPI stuck CPUs for backtrace before trigger_allbutself_cpu_backtrace(),
which does the same thing again.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/powerpc/kernel/watchdog.c | 19 +++++++++++--------
 1 file changed, 11 insertions(+), 8 deletions(-)

diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index 532a1adbe89b..920e61c79f47 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c
@@ -133,15 +133,18 @@ static void watchdog_smp_panic(int cpu, u64 tb)
 	pr_emerg("Watchdog CPU:%d detected Hard LOCKUP other CPUS:%*pbl\n",
 			cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 
-	/*
-	 * Try to trigger the stuck CPUs.
-	 */
-	for_each_cpu(c, &wd_smp_cpus_pending) {
-		if (c == cpu)
-			continue;
-		smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
+	if (!sysctl_hardlockup_all_cpu_backtrace) {
+		/*
+		 * Try to trigger the stuck CPUs, unless we are going to
+		 * get a backtrace on all of them anyway.
+		 */
+		for_each_cpu(c, &wd_smp_cpus_pending) {
+			if (c == cpu)
+				continue;
+			smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
+		}
+		smp_flush_nmi_ipi(1000000);
 	}
-	smp_flush_nmi_ipi(1000000);
 
 	/* Take the stuck CPUs out of the watch group */
 	set_cpumask_stuck(&wd_smp_cpus_pending, tb);
-- 
2.13.3

[PATCH 3/5] powerpc/watchdog: do not trigger SMP crash from touch_nmi_watchdog

[Nicholas Piggin]

In xmon, touch_nmi_watchdog() is not expected to be checking that
other CPUs have not touched the watchdog, so the code will just
call touch_nmi_watchdog() once before re-enabling hard interrupts.

Just update our CPU's state, and ignore apparently stuck SMP threads.

Arguably touch_nmi_watchdog should check for SMP lockups, and
callers should be fixed, but that's not trivial for the input code
of xmon.
---
 arch/powerpc/kernel/watchdog.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index 920e61c79f47..1fb9379dc683 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c
@@ -277,9 +277,12 @@ void arch_touch_nmi_watchdog(void)
 {
 	unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 	int cpu = smp_processor_id();
+	u64 tb = get_tb();
 
-	if (get_tb() - per_cpu(wd_timer_tb, cpu) >= ticks)
-		watchdog_timer_interrupt(cpu);
+	if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
+		per_cpu(wd_timer_tb, cpu) = tb;
+		wd_smp_clear_cpu_pending(cpu, tb);
+	}
 }
 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 
-- 
2.13.3

[PATCH 4/5] powerpc/xmon: avoid tripping SMP hardlockup watchdog

[Nicholas Piggin]

The SMP hardlockup watchdog cross-checks other CPUs for lockups,
which causes xmon headaches because it's assuming interrupts
hard disabled means no watchdog troubles. Try to improve that by
calling touch_nmi_watchdog() in obvious places where secondaries
are spinning.

Also annotate these spin loops with spin_begin/end calls.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/powerpc/xmon/xmon.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/arch/powerpc/xmon/xmon.c b/arch/powerpc/xmon/xmon.c
index 33351c6704b1..d9a12102b111 100644
--- a/arch/powerpc/xmon/xmon.c
+++ b/arch/powerpc/xmon/xmon.c
@@ -530,14 +530,19 @@ static int xmon_core(struct pt_regs *regs, int fromipi)
 
  waiting:
 	secondary = 1;
+	spin_begin();
 	while (secondary && !xmon_gate) {
 		if (in_xmon == 0) {
-			if (fromipi)
+			if (fromipi) {
+				spin_end();
 				goto leave;
+			}
 			secondary = test_and_set_bit(0, &in_xmon);
 		}
-		barrier();
+		spin_cpu_relax();
+		touch_nmi_watchdog();
 	}
+	spin_end();
 
 	if (!secondary && !xmon_gate) {
 		/* we are the first cpu to come in */
@@ -568,21 +573,25 @@ static int xmon_core(struct pt_regs *regs, int fromipi)
 		mb();
 		xmon_gate = 1;
 		barrier();
+		touch_nmi_watchdog();
 	}
 
  cmdloop:
 	while (in_xmon) {
 		if (secondary) {
+			spin_begin();
 			if (cpu == xmon_owner) {
 				if (!test_and_set_bit(0, &xmon_taken)) {
 					secondary = 0;
+					spin_end();
 					continue;
 				}
 				/* missed it */
 				while (cpu == xmon_owner)
-					barrier();
+					spin_cpu_relax();
 			}
-			barrier();
+			spin_cpu_relax();
+			touch_nmi_watchdog();
 		} else {
 			cmd = cmds(regs);
 			if (cmd != 0) {
-- 
2.13.3

[PATCH 5/5] powerpc/powernv: implement NMI IPI with OPAL_SIGNAL_SYSTEM_RESET

[Nicholas Piggin]

This allows MSR[EE]=0 lockups to be detected on an OPAL (bare metal)
system similarly to the hcall NMI IPI on pseries guests, when the
platform/firmware supports it.

This is an example of CPU10 spinning with interrupts hard disabled:

Watchdog CPU:32 detected Hard LOCKUP other CPUS:10
Watchdog CPU:10 Hard LOCKUP
CPU: 10 PID: 4410 Comm: bash Not tainted 4.13.0-rc7-00074-ge89ce1f89f62-dirty #34
task: c0000003a82b4400 task.stack: c0000003af55c000
NIP: c0000000000a7b38 LR: c000000000659044 CTR: c0000000000a7b00
REGS: c00000000fd23d80 TRAP: 0100   Not tainted  (4.13.0-rc7-00074-ge89ce1f89f62-dirty)
MSR: 90000000000c1033 <SF,HV,ME,IR,DR,RI,LE>
CR: 28422222  XER: 20000000
CFAR: c0000000000a7b38 SOFTE: 0
GPR00: c000000000659044 c0000003af55fbb0 c000000001072a00 0000000000000078
GPR04: c0000003c81b5c80 c0000003c81cc7e8 9000000000009033 0000000000000000
GPR08: 0000000000000000 c0000000000a7b00 0000000000000001 9000000000001003
GPR12: c0000000000a7b00 c00000000fd83200 0000000010180df8 0000000010189e60
GPR16: 0000000010189ed8 0000000010151270 000000001018bd88 000000001018de78
GPR20: 00000000370a0668 0000000000000001 00000000101645e0 0000000010163c10
GPR24: 00007fffd14d6294 00007fffd14d6290 c000000000fba6f0 0000000000000004
GPR28: c000000000f351d8 0000000000000078 c000000000f4095c 0000000000000000
NIP [c0000000000a7b38] sysrq_handle_xmon+0x38/0x40
LR [c000000000659044] __handle_sysrq+0xe4/0x270
Call Trace:
[c0000003af55fbd0] [c000000000659044] __handle_sysrq+0xe4/0x270
[c0000003af55fc70] [c000000000659810] write_sysrq_trigger+0x70/0xa0
[c0000003af55fca0] [c0000000003da650] proc_reg_write+0xb0/0x110
[c0000003af55fcf0] [c0000000003423bc] __vfs_write+0x6c/0x1b0
[c0000003af55fd90] [c000000000344398] vfs_write+0xd8/0x240
[c0000003af55fde0] [c00000000034632c] SyS_write+0x6c/0x110
[c0000003af55fe30] [c00000000000b220] system_call+0x58/0x6c

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/powerpc/include/asm/opal-api.h            |  1 +
 arch/powerpc/include/asm/opal.h                |  2 ++
 arch/powerpc/kernel/irq.c                      | 43 +++++++++++++++++++---
 arch/powerpc/platforms/powernv/opal-wrappers.S |  1 +
 arch/powerpc/platforms/powernv/powernv.h       |  1 +
 arch/powerpc/platforms/powernv/setup.c         |  3 ++
 arch/powerpc/platforms/powernv/smp.c           | 50 ++++++++++++++++++++++++++
 7 files changed, 97 insertions(+), 4 deletions(-)

diff --git a/arch/powerpc/include/asm/opal-api.h b/arch/powerpc/include/asm/opal-api.h
index 450a60b81d2a..9d191ebea706 100644
--- a/arch/powerpc/include/asm/opal-api.h
+++ b/arch/powerpc/include/asm/opal-api.h
@@ -188,6 +188,7 @@
 #define OPAL_XIVE_DUMP				142
 #define OPAL_XIVE_RESERVED3			143
 #define OPAL_XIVE_RESERVED4			144
+#define OPAL_SIGNAL_SYSTEM_RESET		145
 #define OPAL_NPU_INIT_CONTEXT			146
 #define OPAL_NPU_DESTROY_CONTEXT		147
 #define OPAL_NPU_MAP_LPAR			148
diff --git a/arch/powerpc/include/asm/opal.h b/arch/powerpc/include/asm/opal.h
index 726c23304a57..7d7613c49f2b 100644
--- a/arch/powerpc/include/asm/opal.h
+++ b/arch/powerpc/include/asm/opal.h
@@ -281,6 +281,8 @@ int opal_get_power_shift_ratio(u32 handle, int token, u32 *psr);
 int opal_set_power_shift_ratio(u32 handle, int token, u32 psr);
 int opal_sensor_group_clear(u32 group_hndl, int token);
 
+int64_t opal_signal_system_reset(int32_t cpu);
+
 /* Internal functions */
 extern int early_init_dt_scan_opal(unsigned long node, const char *uname,
 				   int depth, void *data);
diff --git a/arch/powerpc/kernel/irq.c b/arch/powerpc/kernel/irq.c
index 4e65bf82f5e0..472d294d0df5 100644
--- a/arch/powerpc/kernel/irq.c
+++ b/arch/powerpc/kernel/irq.c
@@ -394,11 +394,21 @@ bool prep_irq_for_idle_irqsoff(void)
 /*
  * Take the SRR1 wakeup reason, index into this table to find the
  * appropriate irq_happened bit.
+ *
+ * Sytem reset exceptions taken in idle state also come through here,
+ * but they are NMI interrupts so do not need to wait for IRQs to be
+ * restored, and should be taken as early as practical. These are marked
+ * with 0xff in the table. The Power ISA specifies 0100b as the system
+ * reset interrupt reason, but POWER9 DD1 can set 0010b.
  */
+#define IRQ_SYSTEM_RESET	0xff
+
 static const u8 srr1_to_lazyirq[0x10] = {
-	0, 0, 0,
-	PACA_IRQ_DBELL,
 	0,
+	0,
+	IRQ_SYSTEM_RESET,
+	PACA_IRQ_DBELL,
+	IRQ_SYSTEM_RESET,
 	PACA_IRQ_DBELL,
 	PACA_IRQ_DEC,
 	0,
@@ -407,15 +417,40 @@ static const u8 srr1_to_lazyirq[0x10] = {
 	PACA_IRQ_HMI,
 	0, 0, 0, 0, 0 };
 
+static noinline void replay_system_reset(void)
+{
+	struct pt_regs regs;
+
+	ppc_save_regs(&regs);
+	regs.trap = 0x100;
+	get_paca()->in_nmi = 1;
+	system_reset_exception(&regs);
+	get_paca()->in_nmi = 0;
+}
+
 void irq_set_pending_from_srr1(unsigned long srr1)
 {
 	unsigned int idx = (srr1 & SRR1_WAKEMASK_P8) >> 18;
+	u8 reason = srr1_to_lazyirq[idx];
+
+	/*
+	 * Take the system reset now, which is immediately after registers
+	 * are restored from idle. It's an NMI, so interrupts need not be
+	 * re-enabled when it's taken.
+	 */
+	if (unlikely(reason == IRQ_SYSTEM_RESET)) {
+		replay_system_reset();
+		return;
+	}
 
 	/*
 	 * The 0 index (SRR1[42:45]=b0000) must always evaluate to 0,
-	 * so this can be called unconditionally with srr1 wake reason.
+	 * so this can be called unconditionally with the SRR1 wake
+	 * reason as returned by the idle code. If a future CPU was to
+	 * designate this as an interrupt reason, then a new index for
+	 * no interrupt must be assigned.
 	 */
-	local_paca->irq_happened |= srr1_to_lazyirq[idx];
+	local_paca->irq_happened |= reason;
 }
 #endif /* CONFIG_PPC_BOOK3S */
 
diff --git a/arch/powerpc/platforms/powernv/opal-wrappers.S b/arch/powerpc/platforms/powernv/opal-wrappers.S
index 8c1ede2d3f7e..37cd170201a2 100644
--- a/arch/powerpc/platforms/powernv/opal-wrappers.S
+++ b/arch/powerpc/platforms/powernv/opal-wrappers.S
@@ -307,6 +307,7 @@ OPAL_CALL(opal_xive_get_vp_info,		OPAL_XIVE_GET_VP_INFO);
 OPAL_CALL(opal_xive_set_vp_info,		OPAL_XIVE_SET_VP_INFO);
 OPAL_CALL(opal_xive_sync,			OPAL_XIVE_SYNC);
 OPAL_CALL(opal_xive_dump,			OPAL_XIVE_DUMP);
+OPAL_CALL(opal_signal_system_reset,		OPAL_SIGNAL_SYSTEM_RESET);
 OPAL_CALL(opal_npu_init_context,		OPAL_NPU_INIT_CONTEXT);
 OPAL_CALL(opal_npu_destroy_context,		OPAL_NPU_DESTROY_CONTEXT);
 OPAL_CALL(opal_npu_map_lpar,			OPAL_NPU_MAP_LPAR);
diff --git a/arch/powerpc/platforms/powernv/powernv.h b/arch/powerpc/platforms/powernv/powernv.h
index a159d48573d7..49add2037e0d 100644
--- a/arch/powerpc/platforms/powernv/powernv.h
+++ b/arch/powerpc/platforms/powernv/powernv.h
@@ -3,6 +3,7 @@
 
 #ifdef CONFIG_SMP
 extern void pnv_smp_init(void);
+extern int pnv_system_reset_exception(struct pt_regs *regs);
 #else
 static inline void pnv_smp_init(void) { }
 #endif
diff --git a/arch/powerpc/platforms/powernv/setup.c b/arch/powerpc/platforms/powernv/setup.c
index 897aa1400eb8..4fdaa1d7c4cd 100644
--- a/arch/powerpc/platforms/powernv/setup.c
+++ b/arch/powerpc/platforms/powernv/setup.c
@@ -282,6 +282,9 @@ static void __init pnv_setup_machdep_opal(void)
 	ppc_md.restart = pnv_restart;
 	pm_power_off = pnv_power_off;
 	ppc_md.halt = pnv_halt;
+#ifdef CONFIG_SMP
+	ppc_md.system_reset_exception = pnv_system_reset_exception;
+#endif
 	ppc_md.machine_check_exception = opal_machine_check;
 	ppc_md.mce_check_early_recovery = opal_mce_check_early_recovery;
 	ppc_md.hmi_exception_early = opal_hmi_exception_early;
diff --git a/arch/powerpc/platforms/powernv/smp.c b/arch/powerpc/platforms/powernv/smp.c
index c17f81e433f7..23da03eda954 100644
--- a/arch/powerpc/platforms/powernv/smp.c
+++ b/arch/powerpc/platforms/powernv/smp.c
@@ -290,6 +290,54 @@ static void __init pnv_smp_probe(void)
 	}
 }
 
+int pnv_system_reset_exception(struct pt_regs *regs)
+{
+	if (smp_handle_nmi_ipi(regs))
+		return 1;
+	return 0;
+}
+
+static int pnv_cause_nmi_ipi(int cpu)
+{
+	int64_t rc;
+
+	if (cpu >= 0) {
+		rc = opal_signal_system_reset(get_hard_smp_processor_id(cpu));
+		if (rc != OPAL_SUCCESS)
+			return 0;
+		return 1;
+
+	} else if (cpu == NMI_IPI_ALL_OTHERS) {
+		bool success = true;
+		int c;
+
+
+		/*
+		 * We do not use broadcasts (yet), because it's not clear
+		 * exactly what semantics Linux wants or the firmware should
+		 * provide.
+		 */
+		for_each_online_cpu(c) {
+			if (c == smp_processor_id())
+				continue;
+
+			rc = opal_signal_system_reset(
+						get_hard_smp_processor_id(c));
+			if (rc != OPAL_SUCCESS)
+				success = false;
+		}
+		if (success)
+			return 1;
+
+		/*
+		 * Caller will fall back to doorbells, which may pick
+		 * up the remainders.
+		 */
+	}
+
+	return 0;
+}
+
 static struct smp_ops_t pnv_smp_ops = {
 	.message_pass	= NULL, /* Use smp_muxed_ipi_message_pass */
 	.cause_ipi	= NULL,	/* Filled at runtime by pnv_smp_probe() */
@@ -308,6 +356,8 @@ static struct smp_ops_t pnv_smp_ops = {
 /* This is called very early during platform setup_arch */
 void __init pnv_smp_init(void)
 {
+	if (opal_check_token(OPAL_SIGNAL_SYSTEM_RESET))
+		pnv_smp_ops.cause_nmi_ipi = pnv_cause_nmi_ipi;
 	smp_ops = &pnv_smp_ops;
 
 #ifdef CONFIG_HOTPLUG_CPU
-- 
2.13.3

Re: [PATCH 0/5] More NMI IPI enablement work

[Michael Ellerman]

Nicholas Piggin <npiggin@gmail.com> writes:

> Hi,
>
> Here is the NMI IPI enablement for powernv, which requires
> supported firmware and hardware:
>
> https://lists.ozlabs.org/pipermail/skiboot/2017-September/009111.html
>
> And some preceeding patches that fix issues I ran into when
> testing it.

Are any of these critical such that I should pick them up for 4.14 or
earlier?

cheers

Re: [PATCH 0/5] More NMI IPI enablement work

[Nicholas Piggin]

On Thu, 21 Sep 2017 21:10:42 +1000
Michael Ellerman <mpe@ellerman.id.au> wrote:

> Nicholas Piggin <npiggin@gmail.com> writes:
> 
> > Hi,
> >
> > Here is the NMI IPI enablement for powernv, which requires
> > supported firmware and hardware:
> >
> > https://lists.ozlabs.org/pipermail/skiboot/2017-September/009111.html
> >
> > And some preceeding patches that fix issues I ran into when
> > testing it.  
> 
> Are any of these critical such that I should pick them up for 4.14 or
> earlier?

Good point, yes. The hardlockup watchdog now does SMP cross-checking,
so it can blow up when you stay in xmon for about 10 seconds since the
new watchdog was merged. We want patches 3-4 for that.

Patch 1 is not really required, I only ran into it when I tried making
smp_send_stop use the NMI IPI, which we are not doing just yet. That
being said, the watchdog code is new, so probably doesn't hurt to put
in 4.14.

Patch 3 makes all-cpus backtraces a bit more concise. Less critical.

The system reset stuff I guess we'll want to backport for distros, so
4.14 might be reasonable not sure.

Actually patch 5 before you merge it I should remove the DD1
workaround too because we're not supporting that in OPAL anymore.

Thanks,
Nick