x86: Add per-cpu stat counter for APIC ICR read tries

In the IPI delivery slow path (NMI delivery) we retry the ICR
read to check for delivery completion a limited number of times.

[ The reason for the limited retries is that some of the places
  where it is used (cpu boot, kdump, etc) IPI delivery might not
  succeed (due to a firmware bug or system crash, for example)
  and in such a case it is better to give up and resume
  execution of other code. ]

This patch adds a new entry to /proc/interrupts, RTR, which
tells user space the number of times we retried the ICR read in
the IPI delivery slow path.

This should give some insight into how well the APIC
message delivery hardware is working - if the counts are way
too large then we are hitting a (very-) slow path way too
often.
Signed-off-by: NFernando Luis Vazquez Cao <fernando@oss.ntt.co.jp>
Cc: Jörn Engel <joern@logfs.org>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Link: http://lkml.kernel.org/n/tip-vzsp20lo2xdzh5f70g0eis2s@git.kernel.org
[ extended the changelog ]
Signed-off-by: NIngo Molnar <mingo@elte.hu>

arch/x86/include/asm/apic.h

@@ -410,6 +410,9 @@ extern int wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip);
 #endif
 
 #ifdef CONFIG_X86_LOCAL_APIC
+
+DECLARE_PER_CPU(unsigned, icr_read_retry_count);
+
 static inline u32 apic_read(u32 reg)
 {
 	return apic->read(reg);

arch/x86/kernel/apic/apic.c

@@ -79,6 +79,11 @@ DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
 EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid);
 
+/*
+ * ICR read retry counter
+ */
+DEFINE_PER_CPU(unsigned, icr_read_retry_count);
+
 #ifdef CONFIG_X86_32
 
 /*
@@ -250,6 +255,7 @@ u32 native_safe_apic_wait_icr_idle(void)
 		send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
 		if (!send_status)
 			break;
+		percpu_inc(icr_read_retry_count);
 		udelay(100);
 	} while (timeout++ < 1000);
 

arch/x86/kernel/irq.c

@@ -74,6 +74,10 @@ int arch_show_interrupts(struct seq_file *p, int prec)
 	for_each_online_cpu(j)
 		seq_printf(p, "%10u ", irq_stats(j)->apic_irq_work_irqs);
 	seq_printf(p, "  IRQ work interrupts\n");
+	seq_printf(p, "%*s: ", prec, "RTR");
+	for_each_online_cpu(j)
+		seq_printf(p, "%10u ", per_cpu(icr_read_retry_count, j));
+	seq_printf(p, "  APIC ICR read retries\n");
 #endif
 	if (x86_platform_ipi_callback) {
 		seq_printf(p, "%*s: ", prec, "PLT");
@@ -136,6 +140,7 @@ u64 arch_irq_stat_cpu(unsigned int cpu)
 	sum += irq_stats(cpu)->irq_spurious_count;
 	sum += irq_stats(cpu)->apic_perf_irqs;
 	sum += irq_stats(cpu)->apic_irq_work_irqs;
+	sum += per_cpu(icr_read_retry_count, cpu);
 #endif
 	if (x86_platform_ipi_callback)
 		sum += irq_stats(cpu)->x86_platform_ipis;