[PATCH 2/3] x86, perf: Add a separate Arch Perfmon v4 PMI handler

[kan.liang@linux.intel.com]

From: Andi Kleen <ak@linux.intel.com>

Implements counter freezing for Arch Perfmon v4 (Skylake and
newer). This allows to speed up the PMI handler by avoiding
unnecessary MSR writes and make it more accurate.

The Arch Perfmon v4 PMI handler is substantially different than
the older PMI handler.

Differences to the old handler:
- It relies on counter freezing, which eliminates several MSR
writes from the PMI handler and lowers the overhead significantly.

It makes the PMI handler more accurate, as all counters get
frozen atomically as soon as any counter overflows. So there is
much less counting of the PMI handler itself.

With the freezing we don't need to disable or enable counters or
PEBS. Only BTS which does not support auto-freezing still needs to
be explicitly managed.

- The PMU acking is done at the end, not the beginning.
This makes it possible to avoid manual enabling/disabling
of the PMU, instead we just rely on the freezing/acking.

- The APIC is acked before reenabling the PMU, which avoids
problems with LBRs occasionally not getting unfreezed on Skylake.

- Looping is only needed to workaround a corner case which several PMIs
are very close to each other. For common cases, the counters are freezed
during PMI handler. It doesn't need to do re-check.

This patch
- Adds code to enable v4 counter freezing
- Fork <=v3 and >=v4 PMI handlers into separate functions.
- Add kernel parameter to disable counter freezing. It took some time to
  debug counter freezing, so in case there are new problems we added an
  option to turn it off. Would not expect this to be used until there
  are new bugs.
- Only for big core. The patch for small core will be posted later
  separately.

Performance:

When profiling a kernel build on Kabylake with different perf options,
measuring the length of all NMI handlers using the nmi handler
trace point:

V3 is without counter freezing.
V4 is with counter freezing.
The value is the average cost of the PMI handler.
(lower is better)

perf options    `           V3(ns) V4(ns)  delta
-c 100000                   1088   894     -18%
-g -c 100000                1862   1646    -12%
--call-graph lbr -c 100000  3649   3367    -8%
--c.g. dwarf -c 100000      2248   1982    -12%

Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
---
 arch/x86/events/intel/core.c     | 122 +++++++++++++++++++++++++++++++++++++++
 arch/x86/events/perf_event.h     |   6 ++
 arch/x86/include/asm/msr-index.h |   1 +
 3 files changed, 129 insertions(+)

diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
index 9b320a5..03778ce 100644
--- a/arch/x86/events/intel/core.c
+++ b/arch/x86/events/intel/core.c
@@ -1995,6 +1995,18 @@ static void intel_pmu_nhm_enable_all(int added)
 	intel_pmu_enable_all(added);
 }
 
+static void enable_counter_freeze(void)
+{
+	update_debugctlmsr(get_debugctlmsr() |
+			DEBUGCTLMSR_FREEZE_PERFMON_ON_PMI);
+}
+
+static void disable_counter_freeze(void)
+{
+	update_debugctlmsr(get_debugctlmsr() &
+			~DEBUGCTLMSR_FREEZE_PERFMON_ON_PMI);
+}
+
 static inline u64 intel_pmu_get_status(void)
 {
 	u64 status;
@@ -2044,6 +2056,14 @@ static void intel_pmu_disable_event(struct perf_event *event)
 	if (unlikely(event->attr.precise_ip))
 		intel_pmu_pebs_disable(event);
 
+	/*
+	 * We could disable freezing here, but doesn't hurt if it's on.
+	 * perf remembers the state, and someone else will likely
+	 * reinitialize.
+	 *
+	 * This avoids an extra MSR write in many situations.
+	 */
+
 	if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
 		intel_pmu_disable_fixed(hwc);
 		return;
@@ -2119,6 +2139,11 @@ static void intel_pmu_enable_event(struct perf_event *event)
 	if (event->attr.exclude_guest)
 		cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx);
 
+	if (x86_pmu.counter_freezing && !cpuc->frozen_enabled) {
+		enable_counter_freeze();
+		cpuc->frozen_enabled = 1;
+	}
+
 	if (unlikely(event_is_checkpointed(event)))
 		cpuc->intel_cp_status |= (1ull << hwc->idx);
 
@@ -2290,6 +2315,87 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status)
 	return handled;
 }
 
+static bool disable_counter_freezing;
+module_param(disable_counter_freezing, bool, 0444);
+MODULE_PARM_DESC(disable_counter_freezing, "Disable counter freezing feature."
+		"The PMI handler will fall back to generic handler."
+		"Default is false (enable counter freezing feature).");
+
+/*
+ * Simplified handler for Arch Perfmon v4:
+ * - We rely on counter freezing/unfreezing to enable/disable the PMU.
+ * This is done automatically on PMU ack.
+ * - Ack the PMU only after the APIC.
+ */
+
+static int intel_pmu_handle_irq_v4(struct pt_regs *regs)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int handled = 0;
+	bool bts = false;
+	u64 status;
+	int pmu_enabled = cpuc->enabled;
+	int loops = 0;
+
+	/* PMU has been disabled because of counter freezing */
+	cpuc->enabled = 0;
+	if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
+		bts = true;
+		intel_bts_disable_local();
+		handled = intel_pmu_drain_bts_buffer();
+		handled += intel_bts_interrupt();
+	}
+	status = intel_pmu_get_status();
+	if (!status)
+		goto done;
+again:
+	intel_pmu_lbr_read();
+	if (++loops > 100) {
+		static bool warned;
+
+		if (!warned) {
+			WARN(1, "perfevents: irq loop stuck!\n");
+			perf_event_print_debug();
+			warned = true;
+		}
+		intel_pmu_reset();
+		goto done;
+	}
+
+
+	handled += handle_pmi_common(regs, status);
+done:
+	/* Ack the PMI in the APIC */
+	apic_write(APIC_LVTPC, APIC_DM_NMI);
+
+	/*
+	 * Ack the PMU late after the APIC.  This avoids bogus
+	 * freezing on Skylake CPUs.  The acking unfreezes the PMU
+	 */
+	if (status) {
+		intel_pmu_ack_status(status);
+	} else {
+		/*
+		 * CPU may issues two PMIs very close to each other.
+		 * When the PMI handler services the first one, the
+		 * GLOBAL_STATUS is already updated to reflect both.
+		 * When it IRETs, the second PMI is immediately
+		 * handled and it sees clear status. At the meantime,
+		 * there may be a third PMI, because the freezing bit
+		 * isn't set since the ack in first PMI handlers.
+		 * Double check if there is more work to be done.
+		 */
+		status = intel_pmu_get_status();
+		if (status)
+			goto again;
+	}
+
+	if (bts)
+		intel_bts_enable_local();
+	cpuc->enabled = pmu_enabled;
+	return handled;
+}
+
 /*
  * This handler is triggered by the local APIC, so the APIC IRQ handling
  * rules apply:
@@ -3432,6 +3538,11 @@ static void intel_pmu_cpu_dying(int cpu)
 	free_excl_cntrs(cpu);
 
 	fini_debug_store_on_cpu(cpu);
+
+	if (cpuc->frozen_enabled) {
+		cpuc->frozen_enabled = 0;
+		disable_counter_freeze();
+	}
 }
 
 static void intel_pmu_sched_task(struct perf_event_context *ctx,
@@ -4325,6 +4436,8 @@ __init int intel_pmu_init(void)
 		x86_pmu.extra_regs = intel_skl_extra_regs;
 		x86_pmu.pebs_aliases = intel_pebs_aliases_skl;
 		x86_pmu.pebs_prec_dist = true;
+		x86_pmu.counter_freezing = disable_counter_freezing ?
+					   false : true;
 		/* all extra regs are per-cpu when HT is on */
 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
 		x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
@@ -4442,6 +4555,15 @@ __init int intel_pmu_init(void)
 		pr_cont("full-width counters, ");
 	}
 
+	/*
+	 * For arch perfmon 4 use counter freezing to avoid
+	 * several MSR accesses in the PMI.
+	 */
+	if (x86_pmu.counter_freezing) {
+		x86_pmu.handle_irq = intel_pmu_handle_irq_v4;
+		pr_cont("counter freezing, ");
+	}
+
 	kfree(to_free);
 	return 0;
 }
diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
index 1562863..f9d153e 100644
--- a/arch/x86/events/perf_event.h
+++ b/arch/x86/events/perf_event.h
@@ -243,6 +243,11 @@ struct cpu_hw_events {
 	int excl_thread_id; /* 0 or 1 */
 
 	/*
+	 * Counter freezing state.
+	 */
+	int				frozen_enabled;
+
+	/*
 	 * AMD specific bits
 	 */
 	struct amd_nb			*amd_nb;
@@ -561,6 +566,7 @@ struct x86_pmu {
 	struct x86_pmu_quirk *quirks;
 	int		perfctr_second_write;
 	bool		late_ack;
+	bool		counter_freezing;
 	u64		(*limit_period)(struct perf_event *event, u64 l);
 
 	/*
diff --git a/arch/x86/include/asm/msr-index.h b/arch/x86/include/asm/msr-index.h
index 68b2c31..4ae4a59 100644
--- a/arch/x86/include/asm/msr-index.h
+++ b/arch/x86/include/asm/msr-index.h
@@ -157,6 +157,7 @@
 #define DEBUGCTLMSR_BTS_OFF_OS		(1UL <<  9)
 #define DEBUGCTLMSR_BTS_OFF_USR		(1UL << 10)
 #define DEBUGCTLMSR_FREEZE_LBRS_ON_PMI	(1UL << 11)
+#define DEBUGCTLMSR_FREEZE_PERFMON_ON_PMI	(1UL << 12)
 #define DEBUGCTLMSR_FREEZE_IN_SMM_BIT	14
 #define DEBUGCTLMSR_FREEZE_IN_SMM	(1UL << DEBUGCTLMSR_FREEZE_IN_SMM_BIT)
 
-- 
2.7.4