On = 30-Jan-2021, at 6:38 PM, Nicholas Piggin <npiggin@gmail.com> = wrote:

This is required in order to allow more significant = differences between
NMI type interrupt handlers and regular = asynchronous handlers.

Signed-off-by: Nicholas Piggin = <npiggin@gmail.com>
---
arch/powerpc/kernel/traps.c = | 31 = +++++++++++++++++++++++++++-
arch/powerpc/perf/core-book3s.c | = 35 ++------------------------------
arch/powerpc/perf/core-fsl-emb.c = | 25 -----------------------
3 files changed, 32 insertions(+), 59 = deletions(-)

diff --git a/arch/powerpc/kernel/traps.c = b/arch/powerpc/kernel/traps.c
index 4349b25807cf..6da3a3642dfb = 100644
--- a/arch/powerpc/kernel/traps.c
+++ = b/arch/powerpc/kernel/traps.c
@@ -1892,11 +1892,40 @@ void = vsx_unavailable_tm(struct pt_regs *regs)
}
#endif /* = CONFIG_PPC_TRANSACTIONAL_MEM */

-void = performance_monitor_exception(struct pt_regs *regs)
+static void = performance_monitor_exception_nmi(struct pt_regs *regs)
+{
+ = nmi_enter();
+
+ = __this_cpu_inc(irq_stat.pmu_irqs);
+
+ = perf_irq(regs);
+
+ nmi_exit();
+}
+
+static = void performance_monitor_exception_async(struct pt_regs = *regs)
{
+ = irq_enter();
+
= __this_cpu_inc(irq_stat.pmu_irqs);

= perf_irq(regs);
+
+ irq_exit();
+}
+
+void = performance_monitor_exception(struct pt_regs *regs)
+{
+ = /*
+ = * On 64-bit, if perf interrupts hit in a = local_irq_disable
+ * (soft-masked) region, we = consider them as NMIs. This is required to
+ * = prevent hash faults on user addresses when reading callchains = (and
+ = * looks better from an irq tracing perspective).
+ = */
+ = if (IS_ENABLED(CONFIG_PPC64) && = unlikely(arch_irq_disabled_regs(regs)))
+ = performance_monitor_exception_nmi(regs);
+ = else
+ = = performance_monitor_exception_async(regs);
}

Hi Nick,
Thanks for the change in tracing of PMI = interrupts.

I tested with this patch that moves perf irq/nmi = handling details into traps.c
and verified the patch works fine for = NMI and asynchronous cases.

Test scenario1: My test kernel module = tries to create one of performance monitor
counter overflow between = local_irq_save/local_irq_restore which should be delivered as an = NMI.

Verified it calls NMI exception handler from my ftrace logs = below

<<>>
dummy_perf = <-performance_monitor_exception_nmi
<<>>

Test = scenario2: My test kernel module tries to create one of performance = monitor counter
overflow between powerpc_local_irq_pmu_save/restore = which should be delivered as an
asynchronous interrupt ( replayed PMI = ).

Verified it calls correct handler from my ftrace = logs:

<<>>
replay_soft_interrupts = <-arch_local_irq_restore
irq_enter = <-performance_monitor_exception_async
irq_enter_rcu = <-performance_monitor_exception_async
dummy_perf = <-performance_monitor_exception_async
irq_exit = <-performance_monitor_exception_async
<<>>

Review= ed-and-Tested-by: Athira Rajeev = <atrajeev@linux.vnet.ibm.com>

#ifdef CONFIG_PPC_ADV_DEBUG_REGS
diff --git = a/arch/powerpc/perf/core-book3s.c = b/arch/powerpc/perf/core-book3s.c
index 28206b1fe172..9fd06010e8b6 = 100644
--- a/arch/powerpc/perf/core-book3s.c
+++ = b/arch/powerpc/perf/core-book3s.c
@@ -110,10 +110,6 @@ static inline = void perf_read_regs(struct pt_regs *regs)
{
= regs->result =3D 0;
}
-static inline int = perf_intr_is_nmi(struct pt_regs *regs)
-{
- return = 0;
-}

static inline int siar_valid(struct pt_regs = *regs)
{
@@ -353,15 +349,6 @@ static inline void = perf_read_regs(struct pt_regs *regs)
regs->result =3D = use_siar;
}

-/*
- * If interrupts were soft-disabled when a = PMU interrupt occurs, treat
- * it as an NMI.
- */
-static = inline int perf_intr_is_nmi(struct pt_regs *regs)
-{
- return = (regs->softe & IRQS_DISABLED);
-}
-
/*
* On = processors like P7+ that have the SIAR-Valid bit, marked = instructions
* must be sampled only if the SIAR-valid bit is = set.
@@ -2279,7 +2266,6 @@ static void __perf_event_interrupt(struct = pt_regs *regs)
struct perf_event = *event;
= unsigned long val[8];
int found, active;
- int = nmi;

= if (cpuhw->n_limited)
freeze_limited_counters(cpuhw, = mfspr(SPRN_PMC5),
@@ -2287,18 +2273,6 @@ static void = __perf_event_interrupt(struct pt_regs *regs)

= perf_read_regs(regs);

- /*
- * If = perf interrupts hit in a local_irq_disable (soft-masked) = region,
- = * we consider them as NMIs. This is required to prevent hash = faults on
- = * user addresses when reading callchains. See the NMI test = in
- = * do_hash_page.
- */
- nmi =3D = perf_intr_is_nmi(regs);
- if (nmi)
- = nmi_enter();
- else
- = irq_enter();
-
/* Read all the PMCs since we'll = need them a bunch of times */
for (i =3D 0; i < = ppmu->n_counter; ++i)
val[i] =3D read_pmc(i + 1);
@@ = -2344,8 +2318,8 @@ static void __perf_event_interrupt(struct pt_regs = *regs)
= = = }
= = }
= }
- = if (!found && !nmi && = printk_ratelimit())
- printk(KERN_WARNING "Can't find = PMC that caused IRQ\n");
+ if (unlikely(!found) && = !arch_irq_disabled_regs(regs))
+ printk_ratelimited(KERN_WARNING = "Can't find PMC that caused IRQ\n");

/*
* Reset = MMCR0 to its normal value. This will set PMXE and
@@ -2355,11 = +2329,6 @@ static void __perf_event_interrupt(struct pt_regs = *regs)
= * we get back out of this interrupt.
= */
= write_mmcr0(cpuhw, cpuhw->mmcr.mmcr0);
-
- if = (nmi)
- = = nmi_exit();
- else
- = irq_exit();
}

static void perf_event_interrupt(struct = pt_regs *regs)
diff --git a/arch/powerpc/perf/core-fsl-emb.c = b/arch/powerpc/perf/core-fsl-emb.c
index e0e7e276bfd2..ee721f420a7b = 100644
--- a/arch/powerpc/perf/core-fsl-emb.c
+++ = b/arch/powerpc/perf/core-fsl-emb.c
@@ -31,19 +31,6 @@ static atomic_t = num_events;
/* Used to avoid races in calling = reserve/release_pmc_hardware */
static = DEFINE_MUTEX(pmc_reserve_mutex);

-/*
- * If interrupts were = soft-disabled when a PMU interrupt occurs, treat
- * it as an = NMI.
- */
-static inline int perf_intr_is_nmi(struct pt_regs = *regs)
-{
-#ifdef __powerpc64__
- return (regs->softe & = IRQS_DISABLED);
-#else
- return = 0;
-#endif
-}
-
static void perf_event_interrupt(struct = pt_regs *regs);

/*
@@ -659,13 +646,6 @@ static void = perf_event_interrupt(struct pt_regs *regs)
struct = perf_event *event;
unsigned long val;
int found = =3D 0;
- = int nmi;
-
- nmi =3D = perf_intr_is_nmi(regs);
- if (nmi)
- = nmi_enter();
- else
- = irq_enter();

for (i =3D 0; i < = ppmu->n_counter; ++i) {
event =3D = cpuhw->event[i];
@@ -690,11 +670,6 @@ static void = perf_event_interrupt(struct pt_regs *regs)
= mtmsr(mfmsr() | MSR_PMM);
mtpmr(PMRN_PMGC0, PMGC0_PMIE | = PMGC0_FCECE);
= isync();
-
- if (nmi)
- = nmi_exit();
- else
- = irq_exit();
}

void hw_perf_event_setup(int cpu)
-- =
2.23.0