x86/nmi: start NMI watchdog on CPU0 after SMP bootstrap

We're noticing a reproducible system boot hang on certain
Skylake platforms where the BIOS is configured in legacy
boot mode with x2APIC disabled. The system stalls immediately
after writing the first SMP initialization sequence into APIC ICR.

The cause of the problem is watchdog NMI handler execution -
somewhere near the end of NMI handling (after it's already
rescheduled the next NMI) it tries to access IO port 0x61
to get the actual NMI reason on CPU0. Unfortunately, this
port is emulated by BIOS using SMIs and this emulation for
some reason takes more time than we expect during INIT-SIPI-SIPI
sequence. As the result, the system is constantly moving between
NMI and SMI handler and not making any progress.

To avoid this, initialize the watchdog after SMP bootstrap on
CPU0 and, additionally, protect the NMI handler by moving
IO port access before NMI re-scheduling. The latter should also
help in case of post boot CPU onlining. Although we're running
watchdog at much lower frequency at this point, it's neveretheless
possible we may trigger the issue anyway.

Signed-off-by: Igor Druzhinin <igor.druzhinin@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: a44f1697968e04fcc6145e3bd51c748b57047240
master date: 2018-02-20 10:16:56 +0100

diff --git a/xen/arch/x86/apic.c b/xen/arch/x86/apic.c
index e6fa2173f0ea602f09a4471dd9c997291bc533bf..b038b170fc39de36453934b46dd280527bdd27f6 100644 (file)
--- a/xen/arch/x86/apic.c
+++ b/xen/arch/x86/apic.c
@@ -688,7 +688,7 @@ void setup_local_APIC(void)
         printk("Leaving ESR disabled.\n");
     }
 
-    if (nmi_watchdog == NMI_LOCAL_APIC)
+    if (nmi_watchdog == NMI_LOCAL_APIC && smp_processor_id())
         setup_apic_nmi_watchdog();
     apic_pm_activate();
 }

diff --git a/xen/arch/x86/smpboot.c b/xen/arch/x86/smpboot.c
index 21a86857399ad2275207c7d657ab6cef210d90ea..535104065a1333bf767430499ad4b8ceeda317ec 100644 (file)
--- a/xen/arch/x86/smpboot.c
+++ b/xen/arch/x86/smpboot.c
@@ -1226,7 +1226,10 @@ int __cpu_up(unsigned int cpu)
 void __init smp_cpus_done(void)
 {
     if ( nmi_watchdog == NMI_LOCAL_APIC )
+    {
+        setup_apic_nmi_watchdog();
         check_nmi_watchdog();
+    }
 
     setup_ioapic_dest();
 

diff --git a/xen/arch/x86/traps.c b/xen/arch/x86/traps.c
index 85424398bcae07d27082f3178aac8c4f69fb7b13..820f94b1eba3fb4540a81224ade5eb82c3f5fa6d 100644 (file)
--- a/xen/arch/x86/traps.c
+++ b/xen/arch/x86/traps.c
@@ -3950,7 +3950,7 @@ static nmi_callback_t *nmi_callback = dummy_nmi_callback;
 void do_nmi(const struct cpu_user_regs *regs)
 {
     unsigned int cpu = smp_processor_id();
-    unsigned char reason;
+    unsigned char reason = 0;
     bool_t handle_unknown = 0;
 
     ++nmi_count(cpu);
@@ -3958,6 +3958,16 @@ void do_nmi(const struct cpu_user_regs *regs)
     if ( nmi_callback(regs, cpu) )
         return;
 
+    /*
+     * Accessing port 0x61 may trap to SMM which has been actually
+     * observed on some production SKX servers. This SMI sometimes
+     * takes enough time for the next NMI tick to happen. By reading
+     * this port before we re-arm the NMI watchdog, we reduce the chance
+     * of having an NMI watchdog expire while in the SMI handler.
+     */
+    if ( cpu == 0 )
+        reason = inb(0x61);
+
     if ( (nmi_watchdog == NMI_NONE) ||
          (!nmi_watchdog_tick(regs) && watchdog_force) )
         handle_unknown = 1;
@@ -3965,7 +3975,6 @@ void do_nmi(const struct cpu_user_regs *regs)
     /* Only the BSP gets external NMIs from the system. */
     if ( cpu == 0 )
     {
-        reason = inb(0x61);
         if ( reason & 0x80 )
             pci_serr_error(regs);
         if ( reason & 0x40 )