tile: support delivering NMIs for multicore backtrace

A new hypervisor service was added some time ago (MDE 4.2.1 or
later, or MDE 4.3 or later) that allows cores to request NMIs
to be delivered to other cores.  Use this facility to deliver
a request that causes a backtrace to be generated on each core,
and hook it into the magic SysRq functionality.

Signed-off-by: Chris Metcalf <cmetcalf@ezchip.com>

arch/tile/kernel/process.c

@@ -27,6 +27,7 @@
 #include <linux/kernel.h>
 #include <linux/tracehook.h>
 #include <linux/signal.h>
+#include <linux/delay.h>
 #include <linux/context_tracking.h>
 #include <asm/stack.h>
 #include <asm/switch_to.h>
@@ -574,3 +575,103 @@ void show_regs(struct pt_regs *regs)
 
 	dump_stack_regs(regs);
 }
+
+/* To ensure stack dump on tiles occurs one by one. */
+static DEFINE_SPINLOCK(backtrace_lock);
+/* To ensure no backtrace occurs before all of the stack dump are done. */
+static atomic_t backtrace_cpus;
+/* The cpu mask to avoid reentrance. */
+static struct cpumask backtrace_mask;
+
+void do_nmi_dump_stack(struct pt_regs *regs)
+{
+	int is_idle = is_idle_task(current) && !in_interrupt();
+	int cpu;
+
+	nmi_enter();
+	cpu = smp_processor_id();
+	if (WARN_ON_ONCE(!cpumask_test_and_clear_cpu(cpu, &backtrace_mask)))
+		goto done;
+
+	spin_lock(&backtrace_lock);
+	if (is_idle)
+		pr_info("CPU: %d idle\n", cpu);
+	else
+		show_regs(regs);
+	spin_unlock(&backtrace_lock);
+	atomic_dec(&backtrace_cpus);
+done:
+	nmi_exit();
+}
+
+#ifdef __tilegx__
+void arch_trigger_all_cpu_backtrace(bool self)
+{
+	struct cpumask mask;
+	HV_Coord tile;
+	unsigned int timeout;
+	int cpu;
+	int ongoing;
+	HV_NMI_Info info[NR_CPUS];
+
+	ongoing = atomic_cmpxchg(&backtrace_cpus, 0, num_online_cpus() - 1);
+	if (ongoing != 0) {
+		pr_err("Trying to do all-cpu backtrace.\n");
+		pr_err("But another all-cpu backtrace is ongoing (%d cpus left)\n",
+		       ongoing);
+		if (self) {
+			pr_err("Reporting the stack on this cpu only.\n");
+			dump_stack();
+		}
+		return;
+	}
+
+	cpumask_copy(&mask, cpu_online_mask);
+	cpumask_clear_cpu(smp_processor_id(), &mask);
+	cpumask_copy(&backtrace_mask, &mask);
+
+	/* Backtrace for myself first. */
+	if (self)
+		dump_stack();
+
+	/* Tentatively dump stack on remote tiles via NMI. */
+	timeout = 100;
+	while (!cpumask_empty(&mask) && timeout) {
+		for_each_cpu(cpu, &mask) {
+			tile.x = cpu_x(cpu);
+			tile.y = cpu_y(cpu);
+			info[cpu] = hv_send_nmi(tile, TILE_NMI_DUMP_STACK, 0);
+			if (info[cpu].result == HV_NMI_RESULT_OK)
+				cpumask_clear_cpu(cpu, &mask);
+		}
+
+		mdelay(10);
+		timeout--;
+	}
+
+	/* Warn about cpus stuck in ICS and decrement their counts here. */
+	if (!cpumask_empty(&mask)) {
+		for_each_cpu(cpu, &mask) {
+			switch (info[cpu].result) {
+			case HV_NMI_RESULT_FAIL_ICS:
+				pr_warn("Skipping stack dump of cpu %d in ICS at pc %#llx\n",
+					cpu, info[cpu].pc);
+				break;
+			case HV_NMI_RESULT_FAIL_HV:
+				pr_warn("Skipping stack dump of cpu %d in hypervisor\n",
+					cpu);
+				break;
+			case HV_ENOSYS:
+				pr_warn("Hypervisor too old to allow remote stack dumps.\n");
+				goto skip_for_each;
+			default:  /* should not happen */
+				pr_warn("Skipping stack dump of cpu %d [%d,%#llx]\n",
+					cpu, info[cpu].result, info[cpu].pc);
+				break;
+			}
+		}
+skip_for_each:
+		atomic_sub(cpumask_weight(&mask), &backtrace_cpus);
+	}
+}
+#endif /* __tilegx_ */