oprofile/x86: fix perfctr nmi reservation for mulitplexing
[safe/jmp/linux-2.6] / arch / x86 / oprofile / nmi_int.c
index fb4902b..7170d1e 100644 (file)
@@ -1,11 +1,14 @@
 /**
  * @file nmi_int.c
  *
- * @remark Copyright 2002-2008 OProfile authors
+ * @remark Copyright 2002-2009 OProfile authors
  * @remark Read the file COPYING
  *
  * @author John Levon <levon@movementarian.org>
  * @author Robert Richter <robert.richter@amd.com>
+ * @author Barry Kasindorf <barry.kasindorf@amd.com>
+ * @author Jason Yeh <jason.yeh@amd.com>
+ * @author Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
  */
 
 #include <linux/init.h>
@@ -16,6 +19,7 @@
 #include <linux/slab.h>
 #include <linux/moduleparam.h>
 #include <linux/kdebug.h>
+#include <linux/cpu.h>
 #include <asm/nmi.h>
 #include <asm/msr.h>
 #include <asm/apic.h>
 #include "op_counter.h"
 #include "op_x86_model.h"
 
-DEFINE_PER_CPU(int, switch_index);
-
-static struct op_x86_model_spec const *model;
+static struct op_x86_model_spec *model;
 static DEFINE_PER_CPU(struct op_msrs, cpu_msrs);
 static DEFINE_PER_CPU(unsigned long, saved_lvtpc);
 
-static int nmi_start(void);
-static void nmi_stop(void);
-static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs);
-static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs);
-static void nmi_cpu_stop(void *dummy);
-static void nmi_cpu_start(void *dummy);
-
 /* 0 == registered but off, 1 == registered and on */
 static int nmi_enabled = 0;
 
-#ifdef CONFIG_PM
+struct op_counter_config counter_config[OP_MAX_COUNTER];
 
-static int nmi_suspend(struct sys_device *dev, pm_message_t state)
+/* common functions */
+
+u64 op_x86_get_ctrl(struct op_x86_model_spec const *model,
+                   struct op_counter_config *counter_config)
 {
-       if (nmi_enabled == 1)
-               nmi_stop();
-       return 0;
+       u64 val = 0;
+       u16 event = (u16)counter_config->event;
+
+       val |= ARCH_PERFMON_EVENTSEL_INT;
+       val |= counter_config->user ? ARCH_PERFMON_EVENTSEL_USR : 0;
+       val |= counter_config->kernel ? ARCH_PERFMON_EVENTSEL_OS : 0;
+       val |= (counter_config->unit_mask & 0xFF) << 8;
+       event &= model->event_mask ? model->event_mask : 0xFF;
+       val |= event & 0xFF;
+       val |= (event & 0x0F00) << 24;
+
+       return val;
 }
 
-static int nmi_resume(struct sys_device *dev)
+
+static int profile_exceptions_notify(struct notifier_block *self,
+                                    unsigned long val, void *data)
 {
-       if (nmi_enabled == 1)
-               nmi_start();
+       struct die_args *args = (struct die_args *)data;
+       int ret = NOTIFY_DONE;
+       int cpu = smp_processor_id();
+
+       switch (val) {
+       case DIE_NMI:
+       case DIE_NMI_IPI:
+               model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu));
+               ret = NOTIFY_STOP;
+               break;
+       default:
+               break;
+       }
+       return ret;
+}
+
+static void nmi_cpu_save_registers(struct op_msrs *msrs)
+{
+       struct op_msr *counters = msrs->counters;
+       struct op_msr *controls = msrs->controls;
+       unsigned int i;
+
+       for (i = 0; i < model->num_counters; ++i) {
+               if (counters[i].addr)
+                       rdmsrl(counters[i].addr, counters[i].saved);
+       }
+
+       for (i = 0; i < model->num_controls; ++i) {
+               if (controls[i].addr)
+                       rdmsrl(controls[i].addr, controls[i].saved);
+       }
+}
+
+static void nmi_cpu_start(void *dummy)
+{
+       struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
+       model->start(msrs);
+}
+
+static int nmi_start(void)
+{
+       on_each_cpu(nmi_cpu_start, NULL, 1);
        return 0;
 }
 
-static struct sysdev_class oprofile_sysclass = {
-       .name           = "oprofile",
-       .resume         = nmi_resume,
-       .suspend        = nmi_suspend,
-};
+static void nmi_cpu_stop(void *dummy)
+{
+       struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
+       model->stop(msrs);
+}
 
-static struct sys_device device_oprofile = {
-       .id     = 0,
-       .cls    = &oprofile_sysclass,
-};
+static void nmi_stop(void)
+{
+       on_each_cpu(nmi_cpu_stop, NULL, 1);
+}
 
-static int __init init_sysfs(void)
+#ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
+
+static DEFINE_PER_CPU(int, switch_index);
+
+static inline int has_mux(void)
 {
-       int error;
+       return !!model->switch_ctrl;
+}
 
-       error = sysdev_class_register(&oprofile_sysclass);
-       if (!error)
-               error = sysdev_register(&device_oprofile);
-       return error;
+inline int op_x86_phys_to_virt(int phys)
+{
+       return __get_cpu_var(switch_index) + phys;
 }
 
-static void exit_sysfs(void)
+inline int op_x86_virt_to_phys(int virt)
 {
-       sysdev_unregister(&device_oprofile);
-       sysdev_class_unregister(&oprofile_sysclass);
+       return virt % model->num_counters;
 }
 
-#else
-#define init_sysfs() do { } while (0)
-#define exit_sysfs() do { } while (0)
-#endif /* CONFIG_PM */
+static void nmi_shutdown_mux(void)
+{
+       int i;
+
+       if (!has_mux())
+               return;
+
+       for_each_possible_cpu(i) {
+               kfree(per_cpu(cpu_msrs, i).multiplex);
+               per_cpu(cpu_msrs, i).multiplex = NULL;
+               per_cpu(switch_index, i) = 0;
+       }
+}
+
+static int nmi_setup_mux(void)
+{
+       size_t multiplex_size =
+               sizeof(struct op_msr) * model->num_virt_counters;
+       int i;
+
+       if (!has_mux())
+               return 1;
+
+       for_each_possible_cpu(i) {
+               per_cpu(cpu_msrs, i).multiplex =
+                       kmalloc(multiplex_size, GFP_KERNEL);
+               if (!per_cpu(cpu_msrs, i).multiplex)
+                       return 0;
+       }
+
+       return 1;
+}
+
+static void nmi_cpu_setup_mux(int cpu, struct op_msrs const * const msrs)
+{
+       int i;
+       struct op_msr *multiplex = msrs->multiplex;
+
+       if (!has_mux())
+               return;
+
+       for (i = 0; i < model->num_virt_counters; ++i) {
+               if (counter_config[i].enabled) {
+                       multiplex[i].saved = -(u64)counter_config[i].count;
+               } else {
+                       multiplex[i].saved = 0;
+               }
+       }
+
+       per_cpu(switch_index, cpu) = 0;
+}
+
+static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs)
+{
+       struct op_msr *counters = msrs->counters;
+       struct op_msr *multiplex = msrs->multiplex;
+       int i;
+
+       for (i = 0; i < model->num_counters; ++i) {
+               int virt = op_x86_phys_to_virt(i);
+               if (counters[i].addr)
+                       rdmsrl(counters[i].addr, multiplex[virt].saved);
+       }
+}
+
+static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs)
+{
+       struct op_msr *counters = msrs->counters;
+       struct op_msr *multiplex = msrs->multiplex;
+       int i;
+
+       for (i = 0; i < model->num_counters; ++i) {
+               int virt = op_x86_phys_to_virt(i);
+               if (counters[i].addr)
+                       wrmsrl(counters[i].addr, multiplex[virt].saved);
+       }
+}
 
 static void nmi_cpu_switch(void *dummy)
 {
@@ -97,17 +222,19 @@ static void nmi_cpu_switch(void *dummy)
        nmi_cpu_save_mpx_registers(msrs);
 
        /* move to next set */
-       si += model->num_hardware_counters;
-       if ((si > model->num_counters) || (counter_config[si].count == 0))
-               per_cpu(switch_index, smp_processor_id()) = 0;
+       si += model->num_counters;
+       if ((si >= model->num_virt_counters) || (counter_config[si].count == 0))
+               per_cpu(switch_index, cpu) = 0;
        else
-               per_cpu(switch_index, smp_processor_id()) = si;
+               per_cpu(switch_index, cpu) = si;
 
+       model->switch_ctrl(model, msrs);
        nmi_cpu_restore_mpx_registers(msrs);
-       model->setup_ctrs(msrs);
+
        nmi_cpu_start(NULL);
 }
 
+
 /*
  * Quick check to see if multiplexing is necessary.
  * The check should be sufficient since counters are used
@@ -115,68 +242,49 @@ static void nmi_cpu_switch(void *dummy)
  */
 static int nmi_multiplex_on(void)
 {
-       return counter_config[model->num_hardware_counters].count ? 0 : -EINVAL;
+       return counter_config[model->num_counters].count ? 0 : -EINVAL;
 }
 
 static int nmi_switch_event(void)
 {
+       if (!has_mux())
+               return -ENOSYS;         /* not implemented */
        if (nmi_multiplex_on() < 0)
-               return -EINVAL;
+               return -EINVAL;         /* not necessary */
 
        on_each_cpu(nmi_cpu_switch, NULL, 1);
 
        return 0;
 }
 
-static int profile_exceptions_notify(struct notifier_block *self,
-                                    unsigned long val, void *data)
+static inline void mux_init(struct oprofile_operations *ops)
 {
-       struct die_args *args = (struct die_args *)data;
-       int ret = NOTIFY_DONE;
-       int cpu = smp_processor_id();
-
-       switch (val) {
-       case DIE_NMI:
-               if (model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu)))
-                       ret = NOTIFY_STOP;
-               break;
-       default:
-               break;
-       }
-       return ret;
+       if (has_mux())
+               ops->switch_events = nmi_switch_event;
 }
 
-static void nmi_cpu_save_registers(struct op_msrs *msrs)
+static void mux_clone(int cpu)
 {
-       unsigned int const nr_ctrs = model->num_counters;
-       unsigned int const nr_ctrls = model->num_controls;
-       struct op_msr *counters = msrs->counters;
-       struct op_msr *controls = msrs->controls;
-       unsigned int i;
+       if (!has_mux())
+               return;
 
-       for (i = 0; i < nr_ctrs; ++i) {
-               if (counters[i].addr) {
-                       rdmsr(counters[i].addr,
-                               counters[i].saved.low,
-                               counters[i].saved.high);
-               }
-       }
-
-       for (i = 0; i < nr_ctrls; ++i) {
-               if (controls[i].addr) {
-                       rdmsr(controls[i].addr,
-                               controls[i].saved.low,
-                               controls[i].saved.high);
-               }
-       }
+       memcpy(per_cpu(cpu_msrs, cpu).multiplex,
+              per_cpu(cpu_msrs, 0).multiplex,
+              sizeof(struct op_msr) * model->num_virt_counters);
 }
 
-static void nmi_save_registers(void *dummy)
-{
-       int cpu = smp_processor_id();
-       struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
-       nmi_cpu_save_registers(msrs);
-}
+#else
+
+inline int op_x86_phys_to_virt(int phys) { return phys; }
+inline int op_x86_virt_to_phys(int virt) { return virt; }
+static inline void nmi_shutdown_mux(void) { }
+static inline int nmi_setup_mux(void) { return 1; }
+static inline void
+nmi_cpu_setup_mux(int cpu, struct op_msrs const * const msrs) { }
+static inline void mux_init(struct oprofile_operations *ops) { }
+static void mux_clone(int cpu) { }
+
+#endif
 
 static void free_msrs(void)
 {
@@ -191,37 +299,32 @@ static void free_msrs(void)
 
 static int allocate_msrs(void)
 {
-       int i, success = 1;
        size_t controls_size = sizeof(struct op_msr) * model->num_controls;
        size_t counters_size = sizeof(struct op_msr) * model->num_counters;
 
+       int i;
        for_each_possible_cpu(i) {
                per_cpu(cpu_msrs, i).counters = kmalloc(counters_size,
-                                                               GFP_KERNEL);
-               if (!per_cpu(cpu_msrs, i).counters) {
-                       success = 0;
-                       break;
-               }
-               per_cpu(cpu_msrs, i).controls =
-                               kmalloc(controls_size, GFP_KERNEL);
-               if (!per_cpu(cpu_msrs, i).controls) {
-                       success = 0;
-                       break;
-               }
+                                                       GFP_KERNEL);
+               if (!per_cpu(cpu_msrs, i).counters)
+                       return 0;
+               per_cpu(cpu_msrs, i).controls = kmalloc(controls_size,
+                                                       GFP_KERNEL);
+               if (!per_cpu(cpu_msrs, i).controls)
+                       return 0;
        }
 
-       if (!success)
-               free_msrs();
-
-       return success;
+       return 1;
 }
 
 static void nmi_cpu_setup(void *dummy)
 {
        int cpu = smp_processor_id();
        struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
+       nmi_cpu_save_registers(msrs);
        spin_lock(&oprofilefs_lock);
-       model->setup_ctrs(msrs);
+       model->setup_ctrs(model, msrs);
+       nmi_cpu_setup_mux(cpu, msrs);
        spin_unlock(&oprofilefs_lock);
        per_cpu(saved_lvtpc, cpu) = apic_read(APIC_LVTPC);
        apic_write(APIC_LVTPC, APIC_DM_NMI);
@@ -230,7 +333,7 @@ static void nmi_cpu_setup(void *dummy)
 static struct notifier_block profile_exceptions_nb = {
        .notifier_call = profile_exceptions_notify,
        .next = NULL,
-       .priority = 0
+       .priority = 2
 };
 
 static int nmi_setup(void)
@@ -239,94 +342,57 @@ static int nmi_setup(void)
        int cpu;
 
        if (!allocate_msrs())
-               return -ENOMEM;
+               err = -ENOMEM;
+       else if (!nmi_setup_mux())
+               err = -ENOMEM;
+       else
+               err = register_die_notifier(&profile_exceptions_nb);
 
-       err = register_die_notifier(&profile_exceptions_nb);
        if (err) {
                free_msrs();
+               nmi_shutdown_mux();
                return err;
        }
 
-       /*
-        * We need to serialize save and setup for HT because the subset
+       /* We need to serialize save and setup for HT because the subset
         * of msrs are distinct for save and setup operations
         */
 
        /* Assume saved/restored counters are the same on all CPUs */
        model->fill_in_addresses(&per_cpu(cpu_msrs, 0));
        for_each_possible_cpu(cpu) {
-               if (cpu != 0) {
-                       memcpy(per_cpu(cpu_msrs, cpu).counters,
-                               per_cpu(cpu_msrs, 0).counters,
-                               sizeof(struct op_msr) * model->num_counters);
-
-                       memcpy(per_cpu(cpu_msrs, cpu).controls,
-                               per_cpu(cpu_msrs, 0).controls,
-                               sizeof(struct op_msr) * model->num_controls);
-               }
-       }
-       on_each_cpu(nmi_save_registers, NULL, 1);
-       on_each_cpu(nmi_cpu_setup, NULL, 1);
-       nmi_enabled = 1;
-       return 0;
-}
+               if (!cpu)
+                       continue;
 
-static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs)
-{
-       unsigned int si = __get_cpu_var(switch_index);
-       unsigned int const nr_ctrs = model->num_hardware_counters;
-       struct op_msr *counters = &msrs->counters[si];
-       unsigned int i;
+               memcpy(per_cpu(cpu_msrs, cpu).counters,
+                      per_cpu(cpu_msrs, 0).counters,
+                      sizeof(struct op_msr) * model->num_counters);
 
-       for (i = 0; i < nr_ctrs; ++i) {
-               int offset = i + si;
-               if (counters[offset].addr) {
-                       rdmsr(counters[offset].addr,
-                               counters[offset].multiplex.low,
-                               counters[offset].multiplex.high);
-               }
-       }
-}
+               memcpy(per_cpu(cpu_msrs, cpu).controls,
+                      per_cpu(cpu_msrs, 0).controls,
+                      sizeof(struct op_msr) * model->num_controls);
 
-static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs)
-{
-       unsigned int si = __get_cpu_var(switch_index);
-       unsigned int const nr_ctrs = model->num_hardware_counters;
-       struct op_msr *counters = &msrs->counters[si];
-       unsigned int i;
-
-       for (i = 0; i < nr_ctrs; ++i) {
-               int offset = i + si;
-               if (counters[offset].addr) {
-                       wrmsr(counters[offset].addr,
-                               counters[offset].multiplex.low,
-                               counters[offset].multiplex.high);
-               }
+               mux_clone(cpu);
        }
+       on_each_cpu(nmi_cpu_setup, NULL, 1);
+       nmi_enabled = 1;
+       return 0;
 }
 
 static void nmi_cpu_restore_registers(struct op_msrs *msrs)
 {
-       unsigned int const nr_ctrs = model->num_counters;
-       unsigned int const nr_ctrls = model->num_controls;
        struct op_msr *counters = msrs->counters;
        struct op_msr *controls = msrs->controls;
        unsigned int i;
 
-       for (i = 0; i < nr_ctrls; ++i) {
-               if (controls[i].addr) {
-                       wrmsr(controls[i].addr,
-                               controls[i].saved.low,
-                               controls[i].saved.high);
-               }
+       for (i = 0; i < model->num_controls; ++i) {
+               if (controls[i].addr)
+                       wrmsrl(controls[i].addr, controls[i].saved);
        }
 
-       for (i = 0; i < nr_ctrs; ++i) {
-               if (counters[i].addr) {
-                       wrmsr(counters[i].addr,
-                               counters[i].saved.low,
-                               counters[i].saved.high);
-               }
+       for (i = 0; i < model->num_counters; ++i) {
+               if (counters[i].addr)
+                       wrmsrl(counters[i].addr, counters[i].saved);
        }
 }
 
@@ -334,7 +400,7 @@ static void nmi_cpu_shutdown(void *dummy)
 {
        unsigned int v;
        int cpu = smp_processor_id();
-       struct op_msrs *msrs = &__get_cpu_var(cpu_msrs);
+       struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
 
        /* restoring APIC_LVTPC can trigger an apic error because the delivery
         * mode and vector nr combination can be illegal. That's by design: on
@@ -346,50 +412,27 @@ static void nmi_cpu_shutdown(void *dummy)
        apic_write(APIC_LVTPC, per_cpu(saved_lvtpc, cpu));
        apic_write(APIC_LVTERR, v);
        nmi_cpu_restore_registers(msrs);
-       __get_cpu_var(switch_index) = 0;
 }
 
 static void nmi_shutdown(void)
 {
-       struct op_msrs *msrs = &get_cpu_var(cpu_msrs);
+       struct op_msrs *msrs;
+
        nmi_enabled = 0;
        on_each_cpu(nmi_cpu_shutdown, NULL, 1);
        unregister_die_notifier(&profile_exceptions_nb);
+       nmi_shutdown_mux();
+       msrs = &get_cpu_var(cpu_msrs);
        model->shutdown(msrs);
        free_msrs();
        put_cpu_var(cpu_msrs);
 }
 
-static void nmi_cpu_start(void *dummy)
-{
-       struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
-       model->start(msrs);
-}
-
-static int nmi_start(void)
-{
-       on_each_cpu(nmi_cpu_start, NULL, 1);
-       return 0;
-}
-
-static void nmi_cpu_stop(void *dummy)
-{
-       struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
-       model->stop(msrs);
-}
-
-static void nmi_stop(void)
-{
-       on_each_cpu(nmi_cpu_stop, NULL, 1);
-}
-
-struct op_counter_config counter_config[OP_MAX_COUNTER];
-
 static int nmi_create_files(struct super_block *sb, struct dentry *root)
 {
        unsigned int i;
 
-       for (i = 0; i < model->num_counters; ++i) {
+       for (i = 0; i < model->num_virt_counters; ++i) {
                struct dentry *dir;
                char buf[4];
 
@@ -398,7 +441,7 @@ static int nmi_create_files(struct super_block *sb, struct dentry *root)
                 * NOTE:  assumes 1:1 mapping here (that counters are organized
                 *        sequentially in their struct assignment).
                 */
-               if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i)))
+               if (!avail_to_resrv_perfctr_nmi_bit(op_x86_virt_to_phys(i)))
                        continue;
 
                snprintf(buf,  sizeof(buf), "%d", i);
@@ -409,20 +452,87 @@ static int nmi_create_files(struct super_block *sb, struct dentry *root)
                oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
                oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
                oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
-               counter_config[i].save_count_low = 0;
        }
 
        return 0;
 }
 
-static int p4force;
-module_param(p4force, int, 0);
+#ifdef CONFIG_SMP
+static int oprofile_cpu_notifier(struct notifier_block *b, unsigned long action,
+                                void *data)
+{
+       int cpu = (unsigned long)data;
+       switch (action) {
+       case CPU_DOWN_FAILED:
+       case CPU_ONLINE:
+               smp_call_function_single(cpu, nmi_cpu_start, NULL, 0);
+               break;
+       case CPU_DOWN_PREPARE:
+               smp_call_function_single(cpu, nmi_cpu_stop, NULL, 1);
+               break;
+       }
+       return NOTIFY_DONE;
+}
+
+static struct notifier_block oprofile_cpu_nb = {
+       .notifier_call = oprofile_cpu_notifier
+};
+#endif
+
+#ifdef CONFIG_PM
+
+static int nmi_suspend(struct sys_device *dev, pm_message_t state)
+{
+       /* Only one CPU left, just stop that one */
+       if (nmi_enabled == 1)
+               nmi_cpu_stop(NULL);
+       return 0;
+}
+
+static int nmi_resume(struct sys_device *dev)
+{
+       if (nmi_enabled == 1)
+               nmi_cpu_start(NULL);
+       return 0;
+}
+
+static struct sysdev_class oprofile_sysclass = {
+       .name           = "oprofile",
+       .resume         = nmi_resume,
+       .suspend        = nmi_suspend,
+};
+
+static struct sys_device device_oprofile = {
+       .id     = 0,
+       .cls    = &oprofile_sysclass,
+};
+
+static int __init init_sysfs(void)
+{
+       int error;
+
+       error = sysdev_class_register(&oprofile_sysclass);
+       if (!error)
+               error = sysdev_register(&device_oprofile);
+       return error;
+}
+
+static void exit_sysfs(void)
+{
+       sysdev_unregister(&device_oprofile);
+       sysdev_class_unregister(&oprofile_sysclass);
+}
+
+#else
+#define init_sysfs() do { } while (0)
+#define exit_sysfs() do { } while (0)
+#endif /* CONFIG_PM */
 
 static int __init p4_init(char **cpu_type)
 {
        __u8 cpu_model = boot_cpu_data.x86_model;
 
-       if (!p4force && (cpu_model > 6 || cpu_model == 5))
+       if (cpu_model > 6 || cpu_model == 5)
                return 0;
 
 #ifndef CONFIG_SMP
@@ -448,9 +558,25 @@ static int __init p4_init(char **cpu_type)
        return 0;
 }
 
+static int force_arch_perfmon;
+static int force_cpu_type(const char *str, struct kernel_param *kp)
+{
+       if (!strcmp(str, "arch_perfmon")) {
+               force_arch_perfmon = 1;
+               printk(KERN_INFO "oprofile: forcing architectural perfmon\n");
+       }
+
+       return 0;
+}
+module_param_call(cpu_type, force_cpu_type, NULL, NULL, 0);
+
 static int __init ppro_init(char **cpu_type)
 {
        __u8 cpu_model = boot_cpu_data.x86_model;
+       struct op_x86_model_spec *spec = &op_ppro_spec; /* default */
+
+       if (force_arch_perfmon && cpu_has_arch_perfmon)
+               return 0;
 
        switch (cpu_model) {
        case 0 ... 2:
@@ -460,29 +586,33 @@ static int __init ppro_init(char **cpu_type)
                *cpu_type = "i386/pii";
                break;
        case 6 ... 8:
+       case 10 ... 11:
                *cpu_type = "i386/piii";
                break;
        case 9:
+       case 13:
                *cpu_type = "i386/p6_mobile";
                break;
-       case 10 ... 13:
-               *cpu_type = "i386/p6";
-               break;
        case 14:
                *cpu_type = "i386/core";
                break;
        case 15: case 23:
                *cpu_type = "i386/core_2";
                break;
+       case 0x2e:
        case 26:
-               *cpu_type = "i386/core_2";
+               spec = &op_arch_perfmon_spec;
+               *cpu_type = "i386/core_i7";
+               break;
+       case 28:
+               *cpu_type = "i386/atom";
                break;
        default:
                /* Unknown */
                return 0;
        }
 
-       model = &op_ppro_spec;
+       model = spec;
        return 1;
 }
 
@@ -493,7 +623,7 @@ int __init op_nmi_init(struct oprofile_operations *ops)
 {
        __u8 vendor = boot_cpu_data.x86_vendor;
        __u8 family = boot_cpu_data.x86;
-       char *cpu_type;
+       char *cpu_type = NULL;
        int ret = 0;
 
        if (!cpu_has_apic)
@@ -504,67 +634,80 @@ int __init op_nmi_init(struct oprofile_operations *ops)
                /* Needs to be at least an Athlon (or hammer in 32bit mode) */
 
                switch (family) {
-               default:
-                       return -ENODEV;
                case 6:
-                       model = &op_amd_spec;
                        cpu_type = "i386/athlon";
                        break;
                case 0xf:
-                       model = &op_amd_spec;
-                       /* Actually it could be i386/hammer too, but give
-                        user space an consistent name. */
+                       /*
+                        * Actually it could be i386/hammer too, but
+                        * give user space an consistent name.
+                        */
                        cpu_type = "x86-64/hammer";
                        break;
                case 0x10:
-                       model = &op_amd_spec;
                        cpu_type = "x86-64/family10";
                        break;
                case 0x11:
-                       model = &op_amd_spec;
                        cpu_type = "x86-64/family11h";
                        break;
+               default:
+                       return -ENODEV;
                }
+               model = &op_amd_spec;
                break;
 
        case X86_VENDOR_INTEL:
                switch (family) {
                        /* Pentium IV */
                case 0xf:
-                       if (!p4_init(&cpu_type))
-                               return -ENODEV;
+                       p4_init(&cpu_type);
                        break;
 
                        /* A P6-class processor */
                case 6:
-                       if (!ppro_init(&cpu_type))
-                               return -ENODEV;
+                       ppro_init(&cpu_type);
                        break;
 
                default:
-                       return -ENODEV;
+                       break;
                }
+
+               if (cpu_type)
+                       break;
+
+               if (!cpu_has_arch_perfmon)
+                       return -ENODEV;
+
+               /* use arch perfmon as fallback */
+               cpu_type = "i386/arch_perfmon";
+               model = &op_arch_perfmon_spec;
                break;
 
        default:
                return -ENODEV;
        }
 
+#ifdef CONFIG_SMP
+       register_cpu_notifier(&oprofile_cpu_nb);
+#endif
        /* default values, can be overwritten by model */
-       __get_cpu_var(switch_index) = 0;
-       ops->create_files = nmi_create_files;
-       ops->setup = nmi_setup;
-       ops->shutdown = nmi_shutdown;
-       ops->start = nmi_start;
-       ops->stop = nmi_stop;
-       ops->cpu_type = cpu_type;
-       ops->switch_events = nmi_switch_event;
+       ops->create_files       = nmi_create_files;
+       ops->setup              = nmi_setup;
+       ops->shutdown           = nmi_shutdown;
+       ops->start              = nmi_start;
+       ops->stop               = nmi_stop;
+       ops->cpu_type           = cpu_type;
 
        if (model->init)
                ret = model->init(ops);
        if (ret)
                return ret;
 
+       if (!model->num_virt_counters)
+               model->num_virt_counters = model->num_counters;
+
+       mux_init(ops);
+
        init_sysfs();
        using_nmi = 1;
        printk(KERN_INFO "oprofile: using NMI interrupt.\n");
@@ -573,8 +716,12 @@ int __init op_nmi_init(struct oprofile_operations *ops)
 
 void op_nmi_exit(void)
 {
-       if (using_nmi)
+       if (using_nmi) {
                exit_sysfs();
+#ifdef CONFIG_SMP
+               unregister_cpu_notifier(&oprofile_cpu_nb);
+#endif
+       }
        if (model->exit)
                model->exit();
 }