Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / arch / x86 / platform / uv / uv_nmi.c
diff --git a/kernel/arch/x86/platform/uv/uv_nmi.c b/kernel/arch/x86/platform/uv/uv_nmi.c
new file mode 100644 (file)
index 0000000..7488caf
--- /dev/null
@@ -0,0 +1,716 @@
+/*
+ * SGI NMI support routines
+ *
+ *  This program is free software; you can redistribute it and/or modify
+ *  it under the terms of the GNU General Public License as published by
+ *  the Free Software Foundation; either version 2 of the License, or
+ *  (at your option) any later version.
+ *
+ *  This program is distributed in the hope that it will be useful,
+ *  but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *  GNU General Public License for more details.
+ *
+ *  You should have received a copy of the GNU General Public License
+ *  along with this program; if not, write to the Free Software
+ *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
+ *
+ *  Copyright (c) 2009-2013 Silicon Graphics, Inc.  All Rights Reserved.
+ *  Copyright (c) Mike Travis
+ */
+
+#include <linux/cpu.h>
+#include <linux/delay.h>
+#include <linux/kdb.h>
+#include <linux/kexec.h>
+#include <linux/kgdb.h>
+#include <linux/module.h>
+#include <linux/nmi.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+
+#include <asm/apic.h>
+#include <asm/current.h>
+#include <asm/kdebug.h>
+#include <asm/local64.h>
+#include <asm/nmi.h>
+#include <asm/traps.h>
+#include <asm/uv/uv.h>
+#include <asm/uv/uv_hub.h>
+#include <asm/uv/uv_mmrs.h>
+
+/*
+ * UV handler for NMI
+ *
+ * Handle system-wide NMI events generated by the global 'power nmi' command.
+ *
+ * Basic operation is to field the NMI interrupt on each cpu and wait
+ * until all cpus have arrived into the nmi handler.  If some cpus do not
+ * make it into the handler, try and force them in with the IPI(NMI) signal.
+ *
+ * We also have to lessen UV Hub MMR accesses as much as possible as this
+ * disrupts the UV Hub's primary mission of directing NumaLink traffic and
+ * can cause system problems to occur.
+ *
+ * To do this we register our primary NMI notifier on the NMI_UNKNOWN
+ * chain.  This reduces the number of false NMI calls when the perf
+ * tools are running which generate an enormous number of NMIs per
+ * second (~4M/s for 1024 cpu threads).  Our secondary NMI handler is
+ * very short as it only checks that if it has been "pinged" with the
+ * IPI(NMI) signal as mentioned above, and does not read the UV Hub's MMR.
+ *
+ */
+
+static struct uv_hub_nmi_s **uv_hub_nmi_list;
+
+DEFINE_PER_CPU(struct uv_cpu_nmi_s, uv_cpu_nmi);
+EXPORT_PER_CPU_SYMBOL_GPL(uv_cpu_nmi);
+
+static unsigned long nmi_mmr;
+static unsigned long nmi_mmr_clear;
+static unsigned long nmi_mmr_pending;
+
+static atomic_t        uv_in_nmi;
+static atomic_t uv_nmi_cpu = ATOMIC_INIT(-1);
+static atomic_t uv_nmi_cpus_in_nmi = ATOMIC_INIT(-1);
+static atomic_t uv_nmi_slave_continue;
+static cpumask_var_t uv_nmi_cpu_mask;
+
+/* Values for uv_nmi_slave_continue */
+#define SLAVE_CLEAR    0
+#define SLAVE_CONTINUE 1
+#define SLAVE_EXIT     2
+
+/*
+ * Default is all stack dumps go to the console and buffer.
+ * Lower level to send to log buffer only.
+ */
+static int uv_nmi_loglevel = CONSOLE_LOGLEVEL_DEFAULT;
+module_param_named(dump_loglevel, uv_nmi_loglevel, int, 0644);
+
+/*
+ * The following values show statistics on how perf events are affecting
+ * this system.
+ */
+static int param_get_local64(char *buffer, const struct kernel_param *kp)
+{
+       return sprintf(buffer, "%lu\n", local64_read((local64_t *)kp->arg));
+}
+
+static int param_set_local64(const char *val, const struct kernel_param *kp)
+{
+       /* clear on any write */
+       local64_set((local64_t *)kp->arg, 0);
+       return 0;
+}
+
+static struct kernel_param_ops param_ops_local64 = {
+       .get = param_get_local64,
+       .set = param_set_local64,
+};
+#define param_check_local64(name, p) __param_check(name, p, local64_t)
+
+static local64_t uv_nmi_count;
+module_param_named(nmi_count, uv_nmi_count, local64, 0644);
+
+static local64_t uv_nmi_misses;
+module_param_named(nmi_misses, uv_nmi_misses, local64, 0644);
+
+static local64_t uv_nmi_ping_count;
+module_param_named(ping_count, uv_nmi_ping_count, local64, 0644);
+
+static local64_t uv_nmi_ping_misses;
+module_param_named(ping_misses, uv_nmi_ping_misses, local64, 0644);
+
+/*
+ * Following values allow tuning for large systems under heavy loading
+ */
+static int uv_nmi_initial_delay = 100;
+module_param_named(initial_delay, uv_nmi_initial_delay, int, 0644);
+
+static int uv_nmi_slave_delay = 100;
+module_param_named(slave_delay, uv_nmi_slave_delay, int, 0644);
+
+static int uv_nmi_loop_delay = 100;
+module_param_named(loop_delay, uv_nmi_loop_delay, int, 0644);
+
+static int uv_nmi_trigger_delay = 10000;
+module_param_named(trigger_delay, uv_nmi_trigger_delay, int, 0644);
+
+static int uv_nmi_wait_count = 100;
+module_param_named(wait_count, uv_nmi_wait_count, int, 0644);
+
+static int uv_nmi_retry_count = 500;
+module_param_named(retry_count, uv_nmi_retry_count, int, 0644);
+
+/*
+ * Valid NMI Actions:
+ *  "dump"     - dump process stack for each cpu
+ *  "ips"      - dump IP info for each cpu
+ *  "kdump"    - do crash dump
+ *  "kdb"      - enter KDB (default)
+ *  "kgdb"     - enter KGDB
+ */
+static char uv_nmi_action[8] = "kdb";
+module_param_string(action, uv_nmi_action, sizeof(uv_nmi_action), 0644);
+
+static inline bool uv_nmi_action_is(const char *action)
+{
+       return (strncmp(uv_nmi_action, action, strlen(action)) == 0);
+}
+
+/* Setup which NMI support is present in system */
+static void uv_nmi_setup_mmrs(void)
+{
+       if (uv_read_local_mmr(UVH_NMI_MMRX_SUPPORTED)) {
+               uv_write_local_mmr(UVH_NMI_MMRX_REQ,
+                                       1UL << UVH_NMI_MMRX_REQ_SHIFT);
+               nmi_mmr = UVH_NMI_MMRX;
+               nmi_mmr_clear = UVH_NMI_MMRX_CLEAR;
+               nmi_mmr_pending = 1UL << UVH_NMI_MMRX_SHIFT;
+               pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMRX_TYPE);
+       } else {
+               nmi_mmr = UVH_NMI_MMR;
+               nmi_mmr_clear = UVH_NMI_MMR_CLEAR;
+               nmi_mmr_pending = 1UL << UVH_NMI_MMR_SHIFT;
+               pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMR_TYPE);
+       }
+}
+
+/* Read NMI MMR and check if NMI flag was set by BMC. */
+static inline int uv_nmi_test_mmr(struct uv_hub_nmi_s *hub_nmi)
+{
+       hub_nmi->nmi_value = uv_read_local_mmr(nmi_mmr);
+       atomic_inc(&hub_nmi->read_mmr_count);
+       return !!(hub_nmi->nmi_value & nmi_mmr_pending);
+}
+
+static inline void uv_local_mmr_clear_nmi(void)
+{
+       uv_write_local_mmr(nmi_mmr_clear, nmi_mmr_pending);
+}
+
+/*
+ * If first cpu in on this hub, set hub_nmi "in_nmi" and "owner" values and
+ * return true.  If first cpu in on the system, set global "in_nmi" flag.
+ */
+static int uv_set_in_nmi(int cpu, struct uv_hub_nmi_s *hub_nmi)
+{
+       int first = atomic_add_unless(&hub_nmi->in_nmi, 1, 1);
+
+       if (first) {
+               atomic_set(&hub_nmi->cpu_owner, cpu);
+               if (atomic_add_unless(&uv_in_nmi, 1, 1))
+                       atomic_set(&uv_nmi_cpu, cpu);
+
+               atomic_inc(&hub_nmi->nmi_count);
+       }
+       return first;
+}
+
+/* Check if this is a system NMI event */
+static int uv_check_nmi(struct uv_hub_nmi_s *hub_nmi)
+{
+       int cpu = smp_processor_id();
+       int nmi = 0;
+
+       local64_inc(&uv_nmi_count);
+       this_cpu_inc(uv_cpu_nmi.queries);
+
+       do {
+               nmi = atomic_read(&hub_nmi->in_nmi);
+               if (nmi)
+                       break;
+
+               if (raw_spin_trylock(&hub_nmi->nmi_lock)) {
+
+                       /* check hub MMR NMI flag */
+                       if (uv_nmi_test_mmr(hub_nmi)) {
+                               uv_set_in_nmi(cpu, hub_nmi);
+                               nmi = 1;
+                               break;
+                       }
+
+                       /* MMR NMI flag is clear */
+                       raw_spin_unlock(&hub_nmi->nmi_lock);
+
+               } else {
+                       /* wait a moment for the hub nmi locker to set flag */
+                       cpu_relax();
+                       udelay(uv_nmi_slave_delay);
+
+                       /* re-check hub in_nmi flag */
+                       nmi = atomic_read(&hub_nmi->in_nmi);
+                       if (nmi)
+                               break;
+               }
+
+               /* check if this BMC missed setting the MMR NMI flag */
+               if (!nmi) {
+                       nmi = atomic_read(&uv_in_nmi);
+                       if (nmi)
+                               uv_set_in_nmi(cpu, hub_nmi);
+               }
+
+       } while (0);
+
+       if (!nmi)
+               local64_inc(&uv_nmi_misses);
+
+       return nmi;
+}
+
+/* Need to reset the NMI MMR register, but only once per hub. */
+static inline void uv_clear_nmi(int cpu)
+{
+       struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
+
+       if (cpu == atomic_read(&hub_nmi->cpu_owner)) {
+               atomic_set(&hub_nmi->cpu_owner, -1);
+               atomic_set(&hub_nmi->in_nmi, 0);
+               uv_local_mmr_clear_nmi();
+               raw_spin_unlock(&hub_nmi->nmi_lock);
+       }
+}
+
+/* Ping non-responding cpus attemping to force them into the NMI handler */
+static void uv_nmi_nr_cpus_ping(void)
+{
+       int cpu;
+
+       for_each_cpu(cpu, uv_nmi_cpu_mask)
+               uv_cpu_nmi_per(cpu).pinging = 1;
+
+       apic->send_IPI_mask(uv_nmi_cpu_mask, APIC_DM_NMI);
+}
+
+/* Clean up flags for cpus that ignored both NMI and ping */
+static void uv_nmi_cleanup_mask(void)
+{
+       int cpu;
+
+       for_each_cpu(cpu, uv_nmi_cpu_mask) {
+               uv_cpu_nmi_per(cpu).pinging =  0;
+               uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_OUT;
+               cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
+       }
+}
+
+/* Loop waiting as cpus enter nmi handler */
+static int uv_nmi_wait_cpus(int first)
+{
+       int i, j, k, n = num_online_cpus();
+       int last_k = 0, waiting = 0;
+
+       if (first) {
+               cpumask_copy(uv_nmi_cpu_mask, cpu_online_mask);
+               k = 0;
+       } else {
+               k = n - cpumask_weight(uv_nmi_cpu_mask);
+       }
+
+       udelay(uv_nmi_initial_delay);
+       for (i = 0; i < uv_nmi_retry_count; i++) {
+               int loop_delay = uv_nmi_loop_delay;
+
+               for_each_cpu(j, uv_nmi_cpu_mask) {
+                       if (uv_cpu_nmi_per(j).state) {
+                               cpumask_clear_cpu(j, uv_nmi_cpu_mask);
+                               if (++k >= n)
+                                       break;
+                       }
+               }
+               if (k >= n) {           /* all in? */
+                       k = n;
+                       break;
+               }
+               if (last_k != k) {      /* abort if no new cpus coming in */
+                       last_k = k;
+                       waiting = 0;
+               } else if (++waiting > uv_nmi_wait_count)
+                       break;
+
+               /* extend delay if waiting only for cpu 0 */
+               if (waiting && (n - k) == 1 &&
+                   cpumask_test_cpu(0, uv_nmi_cpu_mask))
+                       loop_delay *= 100;
+
+               udelay(loop_delay);
+       }
+       atomic_set(&uv_nmi_cpus_in_nmi, k);
+       return n - k;
+}
+
+/* Wait until all slave cpus have entered UV NMI handler */
+static void uv_nmi_wait(int master)
+{
+       /* indicate this cpu is in */
+       this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_IN);
+
+       /* if not the first cpu in (the master), then we are a slave cpu */
+       if (!master)
+               return;
+
+       do {
+               /* wait for all other cpus to gather here */
+               if (!uv_nmi_wait_cpus(1))
+                       break;
+
+               /* if not all made it in, send IPI NMI to them */
+               pr_alert("UV: Sending NMI IPI to %d non-responding CPUs: %*pbl\n",
+                        cpumask_weight(uv_nmi_cpu_mask),
+                        cpumask_pr_args(uv_nmi_cpu_mask));
+
+               uv_nmi_nr_cpus_ping();
+
+               /* if all cpus are in, then done */
+               if (!uv_nmi_wait_cpus(0))
+                       break;
+
+               pr_alert("UV: %d CPUs not in NMI loop: %*pbl\n",
+                        cpumask_weight(uv_nmi_cpu_mask),
+                        cpumask_pr_args(uv_nmi_cpu_mask));
+       } while (0);
+
+       pr_alert("UV: %d of %d CPUs in NMI\n",
+               atomic_read(&uv_nmi_cpus_in_nmi), num_online_cpus());
+}
+
+static void uv_nmi_dump_cpu_ip_hdr(void)
+{
+       printk(KERN_DEFAULT
+               "\nUV: %4s %6s %-32s %s   (Note: PID 0 not listed)\n",
+               "CPU", "PID", "COMMAND", "IP");
+}
+
+static void uv_nmi_dump_cpu_ip(int cpu, struct pt_regs *regs)
+{
+       printk(KERN_DEFAULT "UV: %4d %6d %-32.32s ",
+               cpu, current->pid, current->comm);
+
+       printk_address(regs->ip);
+}
+
+/* Dump this cpu's state */
+static void uv_nmi_dump_state_cpu(int cpu, struct pt_regs *regs)
+{
+       const char *dots = " ................................. ";
+
+       if (uv_nmi_action_is("ips")) {
+               if (cpu == 0)
+                       uv_nmi_dump_cpu_ip_hdr();
+
+               if (current->pid != 0)
+                       uv_nmi_dump_cpu_ip(cpu, regs);
+
+       } else if (uv_nmi_action_is("dump")) {
+               printk(KERN_DEFAULT
+                       "UV:%sNMI process trace for CPU %d\n", dots, cpu);
+               show_regs(regs);
+       }
+       this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_DUMP_DONE);
+}
+
+/* Trigger a slave cpu to dump it's state */
+static void uv_nmi_trigger_dump(int cpu)
+{
+       int retry = uv_nmi_trigger_delay;
+
+       if (uv_cpu_nmi_per(cpu).state != UV_NMI_STATE_IN)
+               return;
+
+       uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP;
+       do {
+               cpu_relax();
+               udelay(10);
+               if (uv_cpu_nmi_per(cpu).state
+                               != UV_NMI_STATE_DUMP)
+                       return;
+       } while (--retry > 0);
+
+       pr_crit("UV: CPU %d stuck in process dump function\n", cpu);
+       uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP_DONE;
+}
+
+/* Wait until all cpus ready to exit */
+static void uv_nmi_sync_exit(int master)
+{
+       atomic_dec(&uv_nmi_cpus_in_nmi);
+       if (master) {
+               while (atomic_read(&uv_nmi_cpus_in_nmi) > 0)
+                       cpu_relax();
+               atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
+       } else {
+               while (atomic_read(&uv_nmi_slave_continue))
+                       cpu_relax();
+       }
+}
+
+/* Walk through cpu list and dump state of each */
+static void uv_nmi_dump_state(int cpu, struct pt_regs *regs, int master)
+{
+       if (master) {
+               int tcpu;
+               int ignored = 0;
+               int saved_console_loglevel = console_loglevel;
+
+               pr_alert("UV: tracing %s for %d CPUs from CPU %d\n",
+                       uv_nmi_action_is("ips") ? "IPs" : "processes",
+                       atomic_read(&uv_nmi_cpus_in_nmi), cpu);
+
+               console_loglevel = uv_nmi_loglevel;
+               atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
+               for_each_online_cpu(tcpu) {
+                       if (cpumask_test_cpu(tcpu, uv_nmi_cpu_mask))
+                               ignored++;
+                       else if (tcpu == cpu)
+                               uv_nmi_dump_state_cpu(tcpu, regs);
+                       else
+                               uv_nmi_trigger_dump(tcpu);
+               }
+               if (ignored)
+                       printk(KERN_DEFAULT "UV: %d CPUs ignored NMI\n",
+                               ignored);
+
+               console_loglevel = saved_console_loglevel;
+               pr_alert("UV: process trace complete\n");
+       } else {
+               while (!atomic_read(&uv_nmi_slave_continue))
+                       cpu_relax();
+               while (this_cpu_read(uv_cpu_nmi.state) != UV_NMI_STATE_DUMP)
+                       cpu_relax();
+               uv_nmi_dump_state_cpu(cpu, regs);
+       }
+       uv_nmi_sync_exit(master);
+}
+
+static void uv_nmi_touch_watchdogs(void)
+{
+       touch_softlockup_watchdog_sync();
+       clocksource_touch_watchdog();
+       rcu_cpu_stall_reset();
+       touch_nmi_watchdog();
+}
+
+#if defined(CONFIG_KEXEC)
+static atomic_t uv_nmi_kexec_failed;
+static void uv_nmi_kdump(int cpu, int master, struct pt_regs *regs)
+{
+       /* Call crash to dump system state */
+       if (master) {
+               pr_emerg("UV: NMI executing crash_kexec on CPU%d\n", cpu);
+               crash_kexec(regs);
+
+               pr_emerg("UV: crash_kexec unexpectedly returned, ");
+               if (!kexec_crash_image) {
+                       pr_cont("crash kernel not loaded\n");
+                       atomic_set(&uv_nmi_kexec_failed, 1);
+                       uv_nmi_sync_exit(1);
+                       return;
+               }
+               pr_cont("kexec busy, stalling cpus while waiting\n");
+       }
+
+       /* If crash exec fails the slaves should return, otherwise stall */
+       while (atomic_read(&uv_nmi_kexec_failed) == 0)
+               mdelay(10);
+
+       /* Crash kernel most likely not loaded, return in an orderly fashion */
+       uv_nmi_sync_exit(0);
+}
+
+#else /* !CONFIG_KEXEC */
+static inline void uv_nmi_kdump(int cpu, int master, struct pt_regs *regs)
+{
+       if (master)
+               pr_err("UV: NMI kdump: KEXEC not supported in this kernel\n");
+}
+#endif /* !CONFIG_KEXEC */
+
+#ifdef CONFIG_KGDB
+#ifdef CONFIG_KGDB_KDB
+static inline int uv_nmi_kdb_reason(void)
+{
+       return KDB_REASON_SYSTEM_NMI;
+}
+#else /* !CONFIG_KGDB_KDB */
+static inline int uv_nmi_kdb_reason(void)
+{
+       /* Insure user is expecting to attach gdb remote */
+       if (uv_nmi_action_is("kgdb"))
+               return 0;
+
+       pr_err("UV: NMI error: KDB is not enabled in this kernel\n");
+       return -1;
+}
+#endif /* CONFIG_KGDB_KDB */
+
+/*
+ * Call KGDB/KDB from NMI handler
+ *
+ * Note that if both KGDB and KDB are configured, then the action of 'kgdb' or
+ * 'kdb' has no affect on which is used.  See the KGDB documention for further
+ * information.
+ */
+static void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
+{
+       if (master) {
+               int reason = uv_nmi_kdb_reason();
+               int ret;
+
+               if (reason < 0)
+                       return;
+
+               /* call KGDB NMI handler as MASTER */
+               ret = kgdb_nmicallin(cpu, X86_TRAP_NMI, regs, reason,
+                               &uv_nmi_slave_continue);
+               if (ret) {
+                       pr_alert("KGDB returned error, is kgdboc set?\n");
+                       atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
+               }
+       } else {
+               /* wait for KGDB signal that it's ready for slaves to enter */
+               int sig;
+
+               do {
+                       cpu_relax();
+                       sig = atomic_read(&uv_nmi_slave_continue);
+               } while (!sig);
+
+               /* call KGDB as slave */
+               if (sig == SLAVE_CONTINUE)
+                       kgdb_nmicallback(cpu, regs);
+       }
+       uv_nmi_sync_exit(master);
+}
+
+#else /* !CONFIG_KGDB */
+static inline void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
+{
+       pr_err("UV: NMI error: KGDB is not enabled in this kernel\n");
+}
+#endif /* !CONFIG_KGDB */
+
+/*
+ * UV NMI handler
+ */
+int uv_handle_nmi(unsigned int reason, struct pt_regs *regs)
+{
+       struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
+       int cpu = smp_processor_id();
+       int master = 0;
+       unsigned long flags;
+
+       local_irq_save(flags);
+
+       /* If not a UV System NMI, ignore */
+       if (!this_cpu_read(uv_cpu_nmi.pinging) && !uv_check_nmi(hub_nmi)) {
+               local_irq_restore(flags);
+               return NMI_DONE;
+       }
+
+       /* Indicate we are the first CPU into the NMI handler */
+       master = (atomic_read(&uv_nmi_cpu) == cpu);
+
+       /* If NMI action is "kdump", then attempt to do it */
+       if (uv_nmi_action_is("kdump"))
+               uv_nmi_kdump(cpu, master, regs);
+
+       /* Pause as all cpus enter the NMI handler */
+       uv_nmi_wait(master);
+
+       /* Dump state of each cpu */
+       if (uv_nmi_action_is("ips") || uv_nmi_action_is("dump"))
+               uv_nmi_dump_state(cpu, regs, master);
+
+       /* Call KGDB/KDB if enabled */
+       else if (uv_nmi_action_is("kdb") || uv_nmi_action_is("kgdb"))
+               uv_call_kgdb_kdb(cpu, regs, master);
+
+       /* Clear per_cpu "in nmi" flag */
+       this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_OUT);
+
+       /* Clear MMR NMI flag on each hub */
+       uv_clear_nmi(cpu);
+
+       /* Clear global flags */
+       if (master) {
+               if (cpumask_weight(uv_nmi_cpu_mask))
+                       uv_nmi_cleanup_mask();
+               atomic_set(&uv_nmi_cpus_in_nmi, -1);
+               atomic_set(&uv_nmi_cpu, -1);
+               atomic_set(&uv_in_nmi, 0);
+       }
+
+       uv_nmi_touch_watchdogs();
+       local_irq_restore(flags);
+
+       return NMI_HANDLED;
+}
+
+/*
+ * NMI handler for pulling in CPUs when perf events are grabbing our NMI
+ */
+static int uv_handle_nmi_ping(unsigned int reason, struct pt_regs *regs)
+{
+       int ret;
+
+       this_cpu_inc(uv_cpu_nmi.queries);
+       if (!this_cpu_read(uv_cpu_nmi.pinging)) {
+               local64_inc(&uv_nmi_ping_misses);
+               return NMI_DONE;
+       }
+
+       this_cpu_inc(uv_cpu_nmi.pings);
+       local64_inc(&uv_nmi_ping_count);
+       ret = uv_handle_nmi(reason, regs);
+       this_cpu_write(uv_cpu_nmi.pinging, 0);
+       return ret;
+}
+
+static void uv_register_nmi_notifier(void)
+{
+       if (register_nmi_handler(NMI_UNKNOWN, uv_handle_nmi, 0, "uv"))
+               pr_warn("UV: NMI handler failed to register\n");
+
+       if (register_nmi_handler(NMI_LOCAL, uv_handle_nmi_ping, 0, "uvping"))
+               pr_warn("UV: PING NMI handler failed to register\n");
+}
+
+void uv_nmi_init(void)
+{
+       unsigned int value;
+
+       /*
+        * Unmask NMI on all cpus
+        */
+       value = apic_read(APIC_LVT1) | APIC_DM_NMI;
+       value &= ~APIC_LVT_MASKED;
+       apic_write(APIC_LVT1, value);
+}
+
+void uv_nmi_setup(void)
+{
+       int size = sizeof(void *) * (1 << NODES_SHIFT);
+       int cpu, nid;
+
+       /* Setup hub nmi info */
+       uv_nmi_setup_mmrs();
+       uv_hub_nmi_list = kzalloc(size, GFP_KERNEL);
+       pr_info("UV: NMI hub list @ 0x%p (%d)\n", uv_hub_nmi_list, size);
+       BUG_ON(!uv_hub_nmi_list);
+       size = sizeof(struct uv_hub_nmi_s);
+       for_each_present_cpu(cpu) {
+               nid = cpu_to_node(cpu);
+               if (uv_hub_nmi_list[nid] == NULL) {
+                       uv_hub_nmi_list[nid] = kzalloc_node(size,
+                                                           GFP_KERNEL, nid);
+                       BUG_ON(!uv_hub_nmi_list[nid]);
+                       raw_spin_lock_init(&(uv_hub_nmi_list[nid]->nmi_lock));
+                       atomic_set(&uv_hub_nmi_list[nid]->cpu_owner, -1);
+               }
+               uv_hub_nmi_per(cpu) = uv_hub_nmi_list[nid];
+       }
+       BUG_ON(!alloc_cpumask_var(&uv_nmi_cpu_mask, GFP_KERNEL));
+       uv_register_nmi_notifier();
+}