Merge branch 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip

Pull perf fixes from Ingo Molnar:
 "This tree contains various perf fixes on the kernel side, plus three
  hw/event-enablement late additions:

   - Intel Memory Bandwidth Monitoring events and handling
   - the AMD Accumulated Power Mechanism reporting facility
   - more IOMMU events

  ... and a final round of perf tooling updates/fixes"

* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (44 commits)
  perf llvm: Use strerror_r instead of the thread unsafe strerror one
  perf llvm: Use realpath to canonicalize paths
  perf tools: Unexport some methods unused outside strbuf.c
  perf probe: No need to use formatting strbuf method
  perf help: Use asprintf instead of adhoc equivalents
  perf tools: Remove unused perf_pathdup, xstrdup functions
  perf tools: Do not include stringify.h from the kernel sources
  tools include: Copy linux/stringify.h from the kernel
  tools lib traceevent: Remove redundant CPU output
  perf tools: Remove needless 'extern' from function prototypes
  perf tools: Simplify die() mechanism
  perf tools: Remove unused DIE_IF macro
  perf script: Remove lots of unused arguments
  perf thread: Rename perf_event__preprocess_sample_addr to thread__resolve
  perf machine: Rename perf_event__preprocess_sample to machine__resolve
  perf tools: Add cpumode to struct perf_sample
  perf tests: Forward the perf_sample in the dwarf unwind test
  perf tools: Remove misplaced __maybe_unused
  perf list: Fix documentation of :ppp
  perf bench numa: Fix assertion for nodes bitfield
  ...

2 files changed, 82 insertions, 38 deletions

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 712570d..de24fbc 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -376,8 +376,11 @@ static void update_perf_cpu_limits(void)
 	u64 tmp = perf_sample_period_ns;
 
 	tmp *= sysctl_perf_cpu_time_max_percent;
-	do_div(tmp, 100);
-	ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
+	tmp = div_u64(tmp, 100);
+	if (!tmp)
+		tmp = 1;
+
+	WRITE_ONCE(perf_sample_allowed_ns, tmp);
 }
 
 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
@@ -409,7 +412,13 @@ int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 	if (ret || !write)
 		return ret;
 
-	update_perf_cpu_limits();
+	if (sysctl_perf_cpu_time_max_percent == 100) {
+		printk(KERN_WARNING
+		       "perf: Dynamic interrupt throttling disabled, can hang your system!\n");
+		WRITE_ONCE(perf_sample_allowed_ns, 0);
+	} else {
+		update_perf_cpu_limits();
+	}
 
 	return 0;
 }
@@ -423,62 +432,68 @@ int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 #define NR_ACCUMULATED_SAMPLES 128
 static DEFINE_PER_CPU(u64, running_sample_length);
 
+static u64 __report_avg;
+static u64 __report_allowed;
+
 static void perf_duration_warn(struct irq_work *w)
 {
-	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
-	u64 avg_local_sample_len;
-	u64 local_samples_len;
-
-	local_samples_len = __this_cpu_read(running_sample_length);
-	avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
-
 	printk_ratelimited(KERN_WARNING
-			"perf interrupt took too long (%lld > %lld), lowering "
-			"kernel.perf_event_max_sample_rate to %d\n",
-			avg_local_sample_len, allowed_ns >> 1,
-			sysctl_perf_event_sample_rate);
+		"perf: interrupt took too long (%lld > %lld), lowering "
+		"kernel.perf_event_max_sample_rate to %d\n",
+		__report_avg, __report_allowed,
+		sysctl_perf_event_sample_rate);
 }
 
 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
 
 void perf_sample_event_took(u64 sample_len_ns)
 {
-	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
-	u64 avg_local_sample_len;
-	u64 local_samples_len;
+	u64 max_len = READ_ONCE(perf_sample_allowed_ns);
+	u64 running_len;
+	u64 avg_len;
+	u32 max;
 
-	if (allowed_ns == 0)
+	if (max_len == 0)
 		return;
 
-	/* decay the counter by 1 average sample */
-	local_samples_len = __this_cpu_read(running_sample_length);
-	local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
-	local_samples_len += sample_len_ns;
-	__this_cpu_write(running_sample_length, local_samples_len);
+	/* Decay the counter by 1 average sample. */
+	running_len = __this_cpu_read(running_sample_length);
+	running_len -= running_len/NR_ACCUMULATED_SAMPLES;
+	running_len += sample_len_ns;
+	__this_cpu_write(running_sample_length, running_len);
 
 	/*
-	 * note: this will be biased artifically low until we have
-	 * seen NR_ACCUMULATED_SAMPLES.  Doing it this way keeps us
+	 * Note: this will be biased artifically low until we have
+	 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
 	 * from having to maintain a count.
 	 */
-	avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
-
-	if (avg_local_sample_len <= allowed_ns)
+	avg_len = running_len/NR_ACCUMULATED_SAMPLES;
+	if (avg_len <= max_len)
 		return;
 
-	if (max_samples_per_tick <= 1)
-		return;
+	__report_avg = avg_len;
+	__report_allowed = max_len;
 
-	max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
-	sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
-	perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
+	/*
+	 * Compute a throttle threshold 25% below the current duration.
+	 */
+	avg_len += avg_len / 4;
+	max = (TICK_NSEC / 100) * sysctl_perf_cpu_time_max_percent;
+	if (avg_len < max)
+		max /= (u32)avg_len;
+	else
+		max = 1;
 
-	update_perf_cpu_limits();
+	WRITE_ONCE(perf_sample_allowed_ns, avg_len);
+	WRITE_ONCE(max_samples_per_tick, max);
+
+	sysctl_perf_event_sample_rate = max * HZ;
+	perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
 
 	if (!irq_work_queue(&perf_duration_work)) {
-		early_printk("perf interrupt took too long (%lld > %lld), lowering "
+		early_printk("perf: interrupt took too long (%lld > %lld), lowering "
 			     "kernel.perf_event_max_sample_rate to %d\n",
-			     avg_local_sample_len, allowed_ns >> 1,
+			     __report_avg, __report_allowed,
 			     sysctl_perf_event_sample_rate);
 	}
 }
@@ -4210,6 +4225,14 @@ static void __perf_event_period(struct perf_event *event,
 	active = (event->state == PERF_EVENT_STATE_ACTIVE);
 	if (active) {
 		perf_pmu_disable(ctx->pmu);
+		/*
+		 * We could be throttled; unthrottle now to avoid the tick
+		 * trying to unthrottle while we already re-started the event.
+		 */
+		if (event->hw.interrupts == MAX_INTERRUPTS) {
+			event->hw.interrupts = 0;
+			perf_log_throttle(event, 1);
+		}
 		event->pmu->stop(event, PERF_EF_UPDATE);
 	}
 
@@ -9426,10 +9449,29 @@ perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
 	switch (action & ~CPU_TASKS_FROZEN) {
 
 	case CPU_UP_PREPARE:
+		/*
+		 * This must be done before the CPU comes alive, because the
+		 * moment we can run tasks we can encounter (software) events.
+		 *
+		 * Specifically, someone can have inherited events on kthreadd
+		 * or a pre-existing worker thread that gets re-bound.
+		 */
 		perf_event_init_cpu(cpu);
 		break;
 
 	case CPU_DOWN_PREPARE:
+		/*
+		 * This must be done before the CPU dies because after that an
+		 * active event might want to IPI the CPU and that'll not work
+		 * so great for dead CPUs.
+		 *
+		 * XXX smp_call_function_single() return -ENXIO without a warn
+		 * so we could possibly deal with this.
+		 *
+		 * This is safe against new events arriving because
+		 * sys_perf_event_open() serializes against hotplug using
+		 * get_online_cpus().
+		 */
 		perf_event_exit_cpu(cpu);
 		break;
 	default:
diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 1faad2cf..c61f0cb 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -746,8 +746,10 @@ struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 
 	rb->user_page = all_buf;
 	rb->data_pages[0] = all_buf + PAGE_SIZE;
-	rb->page_order = ilog2(nr_pages);
-	rb->nr_pages = !!nr_pages;
+	if (nr_pages) {
+		rb->nr_pages = 1;
+		rb->page_order = ilog2(nr_pages);
+	}
 
 	ring_buffer_init(rb, watermark, flags);