[S390] perf: add support for s390x CPU counters
Hendrik Brueckner [Fri, 23 Mar 2012 10:13:06 +0000 (11:13 +0100)]
Add a perf PMU to access the CPU-measurement counter facility CPUM CF.
CPUM CF provides multiple counter sets for measuring generic,
problem-state, and crypto activaties.  Also an extended counter set for
the IBM System z10 and IBM z196 mainframes is available.

Counters from the basic and problem-state counter set are mapped to
generic perf hardware events.  Other counters are accessible through
raw events.

For a list of available counter sets and counters, see:

  - The Load-Program-Parameter and the CPU-Measurement Facilities (SA23-2260)
  - The CPU-Measurement Facility Extended Counters Definition for
    z10 and z196 (SA23-2261)

Reviewed-by: Jan Glauber <jang@linux.vnet.ibm.com>
Signed-off-by: Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

arch/s390/include/asm/cpu_mf.h
arch/s390/include/asm/perf_event.h
arch/s390/kernel/Makefile
arch/s390/kernel/perf_cpum_cf.c [new file with mode: 0644]
arch/s390/kernel/perf_event.c [new file with mode: 0644]

index d91dd38..e49db5d 100644 (file)
@@ -1,12 +1,95 @@
+/*
+ * CPU-measurement facilities
+ *
+ *  Copyright IBM Corp. 2012
+ *  Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
+ *            Jan Glauber <jang@linux.vnet.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License (version 2 only)
+ * as published by the Free Software Foundation.
+ */
 #ifndef _ASM_S390_CPU_MF_H
 #define _ASM_S390_CPU_MF_H
 
-#define CPU_MF_INT_SF_MASK     0xffc00000
-
 #define CPU_MF_INT_SF_IAE      (1 << 31)       /* invalid entry address */
 #define CPU_MF_INT_SF_ISE      (1 << 30)       /* incorrect SDBT entry */
 #define CPU_MF_INT_SF_PRA      (1 << 29)       /* program request alert */
 #define CPU_MF_INT_SF_SACA     (1 << 23)       /* sampler auth. change alert */
 #define CPU_MF_INT_SF_LSDA     (1 << 22)       /* loss of sample data alert */
+#define CPU_MF_INT_CF_CACA     (1 <<  7)       /* counter auth. change alert */
+#define CPU_MF_INT_CF_LCDA     (1 <<  6)       /* loss of counter data alert */
+
+#define CPU_MF_INT_CF_MASK     (CPU_MF_INT_CF_CACA|CPU_MF_INT_CF_LCDA)
+#define CPU_MF_INT_SF_MASK     (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE|   \
+                                CPU_MF_INT_SF_PRA|CPU_MF_INT_SF_SACA|  \
+                                CPU_MF_INT_SF_LSDA)
+
+/* CPU measurement facility support */
+static inline int cpum_cf_avail(void)
+{
+       return MACHINE_HAS_SPP && test_facility(67);
+}
+
+static inline int cpum_sf_avail(void)
+{
+       return MACHINE_HAS_SPP && test_facility(68);
+}
+
+
+struct cpumf_ctr_info {
+       u16   cfvn;
+       u16   auth_ctl;
+       u16   enable_ctl;
+       u16   act_ctl;
+       u16   max_cpu;
+       u16   csvn;
+       u16   max_cg;
+       u16   reserved1;
+       u32   reserved2[12];
+} __packed;
+
+/* Query counter information */
+static inline int qctri(struct cpumf_ctr_info *info)
+{
+       int rc = -EINVAL;
+
+       asm volatile (
+               "0:     .insn   s,0xb28e0000,%1\n"
+               "1:     lhi     %0,0\n"
+               "2:\n"
+               EX_TABLE(1b, 2b)
+               : "+d" (rc), "=Q" (*info));
+       return rc;
+}
+
+/* Load CPU-counter-set controls */
+static inline int lcctl(u64 ctl)
+{
+       int cc;
+
+       asm volatile (
+               "       .insn   s,0xb2840000,%1\n"
+               "       ipm     %0\n"
+               "       srl     %0,28\n"
+               : "=d" (cc) : "m" (ctl) : "cc");
+       return cc;
+}
+
+/* Extract CPU counter */
+static inline int ecctr(u64 ctr, u64 *val)
+{
+       register u64 content asm("4") = 0;
+       int cc;
+
+       asm volatile (
+               "       .insn   rre,0xb2e40000,%0,%2\n"
+               "       ipm     %1\n"
+               "       srl     %1,28\n"
+               : "=d" (content), "=d" (cc) : "d" (ctr) : "cc");
+       if (!cc)
+               *val = content;
+       return cc;
+}
 
-#endif
+#endif /* _ASM_S390_CPU_MF_H */
index 4eb444e..7941968 100644 (file)
@@ -1,8 +1,16 @@
 /*
  * Performance event support - s390 specific definitions.
  *
- * Copyright 2009 Martin Schwidefsky, IBM Corporation.
+ * Copyright IBM Corp. 2009, 2012
+ * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
+ *           Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
  */
 
-/* Empty, just to avoid compiling error */
+#include <asm/cpu_mf.h>
 
+/* CPU-measurement counter facility */
+#define PERF_CPUM_CF_MAX_CTR           160
+
+/* Per-CPU flags for PMU states */
+#define PMU_F_RESERVED                 0x1000
+#define PMU_F_ENABLED                  0x2000
index 16b0b43..884b18a 100644 (file)
@@ -48,6 +48,7 @@ obj-$(CONFIG_DYNAMIC_FTRACE)  += ftrace.o
 obj-$(CONFIG_FUNCTION_GRAPH_TRACER) += ftrace.o
 obj-$(CONFIG_FTRACE_SYSCALLS)  += ftrace.o
 obj-$(CONFIG_CRASH_DUMP)       += crash_dump.o
+obj-$(CONFIG_PERF_EVENTS)      += perf_event.o perf_cpum_cf.o
 
 # Kexec part
 S390_KEXEC_OBJS := machine_kexec.o crash.o
diff --git a/arch/s390/kernel/perf_cpum_cf.c b/arch/s390/kernel/perf_cpum_cf.c
new file mode 100644 (file)
index 0000000..8481ecf
--- /dev/null
@@ -0,0 +1,690 @@
+/*
+ * Performance event support for s390x - CPU-measurement Counter Facility
+ *
+ *  Copyright IBM Corp. 2012
+ *  Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License (version 2 only)
+ * as published by the Free Software Foundation.
+ */
+#define KMSG_COMPONENT "cpum_cf"
+#define pr_fmt(fmt)    KMSG_COMPONENT ": " fmt
+
+#include <linux/kernel.h>
+#include <linux/kernel_stat.h>
+#include <linux/perf_event.h>
+#include <linux/percpu.h>
+#include <linux/notifier.h>
+#include <linux/init.h>
+#include <linux/export.h>
+#include <asm/system.h>
+#include <asm/irq.h>
+#include <asm/cpu_mf.h>
+
+/* CPU-measurement counter facility supports these CPU counter sets:
+ * For CPU counter sets:
+ *    Basic counter set:            0-31
+ *    Problem-state counter set:    32-63
+ *    Crypto-activity counter set:  64-127
+ *    Extented counter set:       128-159
+ */
+enum cpumf_ctr_set {
+       /* CPU counter sets */
+       CPUMF_CTR_SET_BASIC   = 0,
+       CPUMF_CTR_SET_USER    = 1,
+       CPUMF_CTR_SET_CRYPTO  = 2,
+       CPUMF_CTR_SET_EXT     = 3,
+
+       /* Maximum number of counter sets */
+       CPUMF_CTR_SET_MAX,
+};
+
+#define CPUMF_LCCTL_ENABLE_SHIFT    16
+#define CPUMF_LCCTL_ACTCTL_SHIFT     0
+static const u64 cpumf_state_ctl[CPUMF_CTR_SET_MAX] = {
+       [CPUMF_CTR_SET_BASIC]   = 0x02,
+       [CPUMF_CTR_SET_USER]    = 0x04,
+       [CPUMF_CTR_SET_CRYPTO]  = 0x08,
+       [CPUMF_CTR_SET_EXT]     = 0x01,
+};
+
+static void ctr_set_enable(u64 *state, int ctr_set)
+{
+       *state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT;
+}
+static void ctr_set_disable(u64 *state, int ctr_set)
+{
+       *state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT);
+}
+static void ctr_set_start(u64 *state, int ctr_set)
+{
+       *state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT;
+}
+static void ctr_set_stop(u64 *state, int ctr_set)
+{
+       *state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT);
+}
+
+/* Local CPUMF event structure */
+struct cpu_hw_events {
+       struct cpumf_ctr_info   info;
+       atomic_t                ctr_set[CPUMF_CTR_SET_MAX];
+       u64                     state, tx_state;
+       unsigned int            flags;
+};
+static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
+       .ctr_set = {
+               [CPUMF_CTR_SET_BASIC]  = ATOMIC_INIT(0),
+               [CPUMF_CTR_SET_USER]   = ATOMIC_INIT(0),
+               [CPUMF_CTR_SET_CRYPTO] = ATOMIC_INIT(0),
+               [CPUMF_CTR_SET_EXT]    = ATOMIC_INIT(0),
+       },
+       .state = 0,
+       .flags = 0,
+};
+
+static int get_counter_set(u64 event)
+{
+       int set = -1;
+
+       if (event < 32)
+               set = CPUMF_CTR_SET_BASIC;
+       else if (event < 64)
+               set = CPUMF_CTR_SET_USER;
+       else if (event < 128)
+               set = CPUMF_CTR_SET_CRYPTO;
+       else if (event < 160)
+               set = CPUMF_CTR_SET_EXT;
+
+       return set;
+}
+
+static int validate_event(const struct hw_perf_event *hwc)
+{
+       switch (hwc->config_base) {
+       case CPUMF_CTR_SET_BASIC:
+       case CPUMF_CTR_SET_USER:
+       case CPUMF_CTR_SET_CRYPTO:
+       case CPUMF_CTR_SET_EXT:
+               /* check for reserved counters */
+               if ((hwc->config >=  6 && hwc->config <=  31) ||
+                   (hwc->config >= 38 && hwc->config <=  63) ||
+                   (hwc->config >= 80 && hwc->config <= 127))
+                       return -EOPNOTSUPP;
+               break;
+       default:
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int validate_ctr_version(const struct hw_perf_event *hwc)
+{
+       struct cpu_hw_events *cpuhw;
+       int err = 0;
+
+       cpuhw = &get_cpu_var(cpu_hw_events);
+
+       /* check required version for counter sets */
+       switch (hwc->config_base) {
+       case CPUMF_CTR_SET_BASIC:
+       case CPUMF_CTR_SET_USER:
+               if (cpuhw->info.cfvn < 1)
+                       err = -EOPNOTSUPP;
+               break;
+       case CPUMF_CTR_SET_CRYPTO:
+       case CPUMF_CTR_SET_EXT:
+               if (cpuhw->info.csvn < 1)
+                       err = -EOPNOTSUPP;
+               break;
+       }
+
+       put_cpu_var(cpu_hw_events);
+       return err;
+}
+
+static int validate_ctr_auth(const struct hw_perf_event *hwc)
+{
+       struct cpu_hw_events *cpuhw;
+       u64 ctrs_state;
+       int err = 0;
+
+       cpuhw = &get_cpu_var(cpu_hw_events);
+
+       /* check authorization for cpu counter sets */
+       ctrs_state = cpumf_state_ctl[hwc->config_base];
+       if (!(ctrs_state & cpuhw->info.auth_ctl))
+               err = -EPERM;
+
+       put_cpu_var(cpu_hw_events);
+       return err;
+}
+
+/*
+ * Change the CPUMF state to active.
+ * Enable and activate the CPU-counter sets according
+ * to the per-cpu control state.
+ */
+static void cpumf_pmu_enable(struct pmu *pmu)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+       int err;
+
+       if (cpuhw->flags & PMU_F_ENABLED)
+               return;
+
+       err = lcctl(cpuhw->state);
+       if (err) {
+               pr_err("Enabling the performance measuring unit "
+                      "failed with rc=%lx\n", err);
+               return;
+       }
+
+       cpuhw->flags |= PMU_F_ENABLED;
+}
+
+/*
+ * Change the CPUMF state to inactive.
+ * Disable and enable (inactive) the CPU-counter sets according
+ * to the per-cpu control state.
+ */
+static void cpumf_pmu_disable(struct pmu *pmu)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+       int err;
+       u64 inactive;
+
+       if (!(cpuhw->flags & PMU_F_ENABLED))
+               return;
+
+       inactive = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
+       err = lcctl(inactive);
+       if (err) {
+               pr_err("Disabling the performance measuring unit "
+                      "failed with rc=%lx\n", err);
+               return;
+       }
+
+       cpuhw->flags &= ~PMU_F_ENABLED;
+}
+
+
+/* Number of perf events counting hardware events */
+static atomic_t num_events = ATOMIC_INIT(0);
+/* Used to avoid races in calling reserve/release_cpumf_hardware */
+static DEFINE_MUTEX(pmc_reserve_mutex);
+
+/* CPU-measurement alerts for the counter facility */
+static void cpumf_measurement_alert(struct ext_code ext_code,
+                                   unsigned int alert, unsigned long unused)
+{
+       struct cpu_hw_events *cpuhw;
+
+       if (!(alert & CPU_MF_INT_CF_MASK))
+               return;
+
+       kstat_cpu(smp_processor_id()).irqs[EXTINT_CPM]++;
+       cpuhw = &__get_cpu_var(cpu_hw_events);
+
+       /* Measurement alerts are shared and might happen when the PMU
+        * is not reserved.  Ignore these alerts in this case. */
+       if (!(cpuhw->flags & PMU_F_RESERVED))
+               return;
+
+       /* counter authorization change alert */
+       if (alert & CPU_MF_INT_CF_CACA)
+               qctri(&cpuhw->info);
+
+       /* loss of counter data alert */
+       if (alert & CPU_MF_INT_CF_LCDA)
+               pr_err("CPU[%i] Counter data was lost\n", smp_processor_id());
+}
+
+#define PMC_INIT      0
+#define PMC_RELEASE   1
+static void setup_pmc_cpu(void *flags)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+
+       switch (*((int *) flags)) {
+       case PMC_INIT:
+               memset(&cpuhw->info, 0, sizeof(cpuhw->info));
+               qctri(&cpuhw->info);
+               cpuhw->flags |= PMU_F_RESERVED;
+               break;
+
+       case PMC_RELEASE:
+               cpuhw->flags &= ~PMU_F_RESERVED;
+               break;
+       }
+
+       /* Disable CPU counter sets */
+       lcctl(0);
+}
+
+/* Initialize the CPU-measurement facility */
+static int reserve_pmc_hardware(void)
+{
+       int flags = PMC_INIT;
+
+       on_each_cpu(setup_pmc_cpu, &flags, 1);
+       measurement_alert_subclass_register();
+
+       return 0;
+}
+
+/* Release the CPU-measurement facility */
+static void release_pmc_hardware(void)
+{
+       int flags = PMC_RELEASE;
+
+       on_each_cpu(setup_pmc_cpu, &flags, 1);
+       measurement_alert_subclass_unregister();
+}
+
+/* Release the PMU if event is the last perf event */
+static void hw_perf_event_destroy(struct perf_event *event)
+{
+       if (!atomic_add_unless(&num_events, -1, 1)) {
+               mutex_lock(&pmc_reserve_mutex);
+               if (atomic_dec_return(&num_events) == 0)
+                       release_pmc_hardware();
+               mutex_unlock(&pmc_reserve_mutex);
+       }
+}
+
+/* CPUMF <-> perf event mappings for kernel+userspace (basic set) */
+static const int cpumf_generic_events_basic[] = {
+       [PERF_COUNT_HW_CPU_CYCLES]          = 0,
+       [PERF_COUNT_HW_INSTRUCTIONS]        = 1,
+       [PERF_COUNT_HW_CACHE_REFERENCES]    = -1,
+       [PERF_COUNT_HW_CACHE_MISSES]        = -1,
+       [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
+       [PERF_COUNT_HW_BRANCH_MISSES]       = -1,
+       [PERF_COUNT_HW_BUS_CYCLES]          = -1,
+};
+/* CPUMF <-> perf event mappings for userspace (problem-state set) */
+static const int cpumf_generic_events_user[] = {
+       [PERF_COUNT_HW_CPU_CYCLES]          = 32,
+       [PERF_COUNT_HW_INSTRUCTIONS]        = 33,
+       [PERF_COUNT_HW_CACHE_REFERENCES]    = -1,
+       [PERF_COUNT_HW_CACHE_MISSES]        = -1,
+       [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
+       [PERF_COUNT_HW_BRANCH_MISSES]       = -1,
+       [PERF_COUNT_HW_BUS_CYCLES]          = -1,
+};
+
+static int __hw_perf_event_init(struct perf_event *event)
+{
+       struct perf_event_attr *attr = &event->attr;
+       struct hw_perf_event *hwc = &event->hw;
+       int err;
+       u64 ev;
+
+       switch (attr->type) {
+       case PERF_TYPE_RAW:
+               /* Raw events are used to access counters directly,
+                * hence do not permit excludes */
+               if (attr->exclude_kernel || attr->exclude_user ||
+                   attr->exclude_hv)
+                       return -EOPNOTSUPP;
+               ev = attr->config;
+               break;
+
+       case PERF_TYPE_HARDWARE:
+               ev = attr->config;
+               /* Count user space (problem-state) only */
+               if (!attr->exclude_user && attr->exclude_kernel) {
+                       if (ev >= ARRAY_SIZE(cpumf_generic_events_user))
+                               return -EOPNOTSUPP;
+                       ev = cpumf_generic_events_user[ev];
+
+               /* No support for kernel space counters only */
+               } else if (!attr->exclude_kernel && attr->exclude_user) {
+                       return -EOPNOTSUPP;
+
+               /* Count user and kernel space */
+               } else {
+                       if (ev >= ARRAY_SIZE(cpumf_generic_events_basic))
+                               return -EOPNOTSUPP;
+                       ev = cpumf_generic_events_basic[ev];
+               }
+               break;
+
+       default:
+               return -ENOENT;
+       }
+
+       if (ev == -1)
+               return -ENOENT;
+
+       if (ev >= PERF_CPUM_CF_MAX_CTR)
+               return -EINVAL;
+
+       /* The CPU measurement counter facility does not have any interrupts
+        * to do sampling.  Sampling must be provided by external means,
+        * for example, by timers.
+        */
+       if (hwc->sample_period)
+               return -EINVAL;
+
+       /* Use the hardware perf event structure to store the counter number
+        * in 'config' member and the counter set to which the counter belongs
+        * in the 'config_base'.  The counter set (config_base) is then used
+        * to enable/disable the counters.
+        */
+       hwc->config = ev;
+       hwc->config_base = get_counter_set(ev);
+
+       /* Validate the counter that is assigned to this event.
+        * Because the counter facility can use numerous counters at the
+        * same time without constraints, it is not necessary to explicity
+        * validate event groups (event->group_leader != event).
+        */
+       err = validate_event(hwc);
+       if (err)
+               return err;
+
+       /* Initialize for using the CPU-measurement counter facility */
+       if (!atomic_inc_not_zero(&num_events)) {
+               mutex_lock(&pmc_reserve_mutex);
+               if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
+                       err = -EBUSY;
+               else
+                       atomic_inc(&num_events);
+               mutex_unlock(&pmc_reserve_mutex);
+       }
+       event->destroy = hw_perf_event_destroy;
+
+       /* Finally, validate version and authorization of the counter set */
+       err = validate_ctr_auth(hwc);
+       if (!err)
+               err = validate_ctr_version(hwc);
+
+       return err;
+}
+
+static int cpumf_pmu_event_init(struct perf_event *event)
+{
+       int err;
+
+       switch (event->attr.type) {
+       case PERF_TYPE_HARDWARE:
+       case PERF_TYPE_HW_CACHE:
+       case PERF_TYPE_RAW:
+               err = __hw_perf_event_init(event);
+               break;
+       default:
+               return -ENOENT;
+       }
+
+       if (unlikely(err) && event->destroy)
+               event->destroy(event);
+
+       return err;
+}
+
+static int hw_perf_event_reset(struct perf_event *event)
+{
+       u64 prev, new;
+       int err;
+
+       do {
+               prev = local64_read(&event->hw.prev_count);
+               err = ecctr(event->hw.config, &new);
+               if (err) {
+                       if (err != 3)
+                               break;
+                       /* The counter is not (yet) available. This
+                        * might happen if the counter set to which
+                        * this counter belongs is in the disabled
+                        * state.
+                        */
+                       new = 0;
+               }
+       } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
+
+       return err;
+}
+
+static int hw_perf_event_update(struct perf_event *event)
+{
+       u64 prev, new, delta;
+       int err;
+
+       do {
+               prev = local64_read(&event->hw.prev_count);
+               err = ecctr(event->hw.config, &new);
+               if (err)
+                       goto out;
+       } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
+
+       delta = (prev <= new) ? new - prev
+                             : (-1ULL - prev) + new + 1;        /* overflow */
+       local64_add(delta, &event->count);
+out:
+       return err;
+}
+
+static void cpumf_pmu_read(struct perf_event *event)
+{
+       if (event->hw.state & PERF_HES_STOPPED)
+               return;
+
+       hw_perf_event_update(event);
+}
+
+static void cpumf_pmu_start(struct perf_event *event, int flags)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+       struct hw_perf_event *hwc = &event->hw;
+
+       if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
+               return;
+
+       if (WARN_ON_ONCE(hwc->config == -1))
+               return;
+
+       if (flags & PERF_EF_RELOAD)
+               WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
+
+       hwc->state = 0;
+
+       /* (Re-)enable and activate the counter set */
+       ctr_set_enable(&cpuhw->state, hwc->config_base);
+       ctr_set_start(&cpuhw->state, hwc->config_base);
+
+       /* The counter set to which this counter belongs can be already active.
+        * Because all counters in a set are active, the event->hw.prev_count
+        * needs to be synchronized.  At this point, the counter set can be in
+        * the inactive or disabled state.
+        */
+       hw_perf_event_reset(event);
+
+       /* increment refcount for this counter set */
+       atomic_inc(&cpuhw->ctr_set[hwc->config_base]);
+}
+
+static void cpumf_pmu_stop(struct perf_event *event, int flags)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+       struct hw_perf_event *hwc = &event->hw;
+
+       if (!(hwc->state & PERF_HES_STOPPED)) {
+               /* Decrement reference count for this counter set and if this
+                * is the last used counter in the set, clear activation
+                * control and set the counter set state to inactive.
+                */
+               if (!atomic_dec_return(&cpuhw->ctr_set[hwc->config_base]))
+                       ctr_set_stop(&cpuhw->state, hwc->config_base);
+               event->hw.state |= PERF_HES_STOPPED;
+       }
+
+       if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
+               hw_perf_event_update(event);
+               event->hw.state |= PERF_HES_UPTODATE;
+       }
+}
+
+static int cpumf_pmu_add(struct perf_event *event, int flags)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+
+       /* Check authorization for the counter set to which this
+        * counter belongs.
+        * For group events transaction, the authorization check is
+        * done in cpumf_pmu_commit_txn().
+        */
+       if (!(cpuhw->flags & PERF_EVENT_TXN))
+               if (validate_ctr_auth(&event->hw))
+                       return -EPERM;
+
+       ctr_set_enable(&cpuhw->state, event->hw.config_base);
+       event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
+
+       if (flags & PERF_EF_START)
+               cpumf_pmu_start(event, PERF_EF_RELOAD);
+
+       perf_event_update_userpage(event);
+
+       return 0;
+}
+
+static void cpumf_pmu_del(struct perf_event *event, int flags)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+
+       cpumf_pmu_stop(event, PERF_EF_UPDATE);
+
+       /* Check if any counter in the counter set is still used.  If not used,
+        * change the counter set to the disabled state.  This also clears the
+        * content of all counters in the set.
+        *
+        * When a new perf event has been added but not yet started, this can
+        * clear enable control and resets all counters in a set.  Therefore,
+        * cpumf_pmu_start() always has to reenable a counter set.
+        */
+       if (!atomic_read(&cpuhw->ctr_set[event->hw.config_base]))
+               ctr_set_disable(&cpuhw->state, event->hw.config_base);
+
+       perf_event_update_userpage(event);
+}
+
+/*
+ * Start group events scheduling transaction.
+ * Set flags to perform a single test at commit time.
+ */
+static void cpumf_pmu_start_txn(struct pmu *pmu)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+
+       perf_pmu_disable(pmu);
+       cpuhw->flags |= PERF_EVENT_TXN;
+       cpuhw->tx_state = cpuhw->state;
+}
+
+/*
+ * Stop and cancel a group events scheduling tranctions.
+ * Assumes cpumf_pmu_del() is called for each successful added
+ * cpumf_pmu_add() during the transaction.
+ */
+static void cpumf_pmu_cancel_txn(struct pmu *pmu)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+
+       WARN_ON(cpuhw->tx_state != cpuhw->state);
+
+       cpuhw->flags &= ~PERF_EVENT_TXN;
+       perf_pmu_enable(pmu);
+}
+
+/*
+ * Commit the group events scheduling transaction.  On success, the
+ * transaction is closed.   On error, the transaction is kept open
+ * until cpumf_pmu_cancel_txn() is called.
+ */
+static int cpumf_pmu_commit_txn(struct pmu *pmu)
+{
+       struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
+       u64 state;
+
+       /* check if the updated state can be scheduled */
+       state = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
+       state >>= CPUMF_LCCTL_ENABLE_SHIFT;
+       if ((state & cpuhw->info.auth_ctl) != state)
+               return -EPERM;
+
+       cpuhw->flags &= ~PERF_EVENT_TXN;
+       perf_pmu_enable(pmu);
+       return 0;
+}
+
+/* Performance monitoring unit for s390x */
+static struct pmu cpumf_pmu = {
+       .pmu_enable   = cpumf_pmu_enable,
+       .pmu_disable  = cpumf_pmu_disable,
+       .event_init   = cpumf_pmu_event_init,
+       .add          = cpumf_pmu_add,
+       .del          = cpumf_pmu_del,
+       .start        = cpumf_pmu_start,
+       .stop         = cpumf_pmu_stop,
+       .read         = cpumf_pmu_read,
+       .start_txn    = cpumf_pmu_start_txn,
+       .commit_txn   = cpumf_pmu_commit_txn,
+       .cancel_txn   = cpumf_pmu_cancel_txn,
+};
+
+static int __cpuinit cpumf_pmu_notifier(struct notifier_block *self,
+                                       unsigned long action, void *hcpu)
+{
+       unsigned int cpu = (long) hcpu;
+       int flags;
+
+       switch (action & ~CPU_TASKS_FROZEN) {
+       case CPU_ONLINE:
+               flags = PMC_INIT;
+               smp_call_function_single(cpu, setup_pmc_cpu, &flags, 1);
+               break;
+       case CPU_DOWN_PREPARE:
+               flags = PMC_RELEASE;
+               smp_call_function_single(cpu, setup_pmc_cpu, &flags, 1);
+               break;
+       default:
+               break;
+       }
+
+       return NOTIFY_OK;
+}
+
+static int __init cpumf_pmu_init(void)
+{
+       int rc;
+
+       if (!cpum_cf_avail())
+               return -ENODEV;
+
+       /* clear bit 15 of cr0 to unauthorize problem-state to
+        * extract measurement counters */
+       ctl_clear_bit(0, 48);
+
+       /* register handler for measurement-alert interruptions */
+       rc = register_external_interrupt(0x1407, cpumf_measurement_alert);
+       if (rc) {
+               pr_err("Registering for CPU-measurement alerts "
+                      "failed with rc=%i\n", rc);
+               goto out;
+       }
+
+       rc = perf_pmu_register(&cpumf_pmu, "cpum_cf", PERF_TYPE_RAW);
+       if (rc) {
+               pr_err("Registering the cpum_cf PMU failed with rc=%i\n", rc);
+               unregister_external_interrupt(0x1407, cpumf_measurement_alert);
+               goto out;
+       }
+       perf_cpu_notifier(cpumf_pmu_notifier);
+out:
+       return rc;
+}
+early_initcall(cpumf_pmu_init);
diff --git a/arch/s390/kernel/perf_event.c b/arch/s390/kernel/perf_event.c
new file mode 100644 (file)
index 0000000..609f985
--- /dev/null
@@ -0,0 +1,125 @@
+/*
+ * Performance event support for s390x
+ *
+ *  Copyright IBM Corp. 2012
+ *  Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License (version 2 only)
+ * as published by the Free Software Foundation.
+ */
+#define KMSG_COMPONENT "perf"
+#define pr_fmt(fmt)    KMSG_COMPONENT ": " fmt
+
+#include <linux/kernel.h>
+#include <linux/perf_event.h>
+#include <linux/percpu.h>
+#include <linux/export.h>
+#include <asm/system.h>
+#include <asm/irq.h>
+#include <asm/cpu_mf.h>
+#include <asm/lowcore.h>
+#include <asm/processor.h>
+
+const char *perf_pmu_name(void)
+{
+       if (cpum_cf_avail() || cpum_sf_avail())
+               return "CPU-measurement facilities (CPUMF)";
+       return "pmu";
+}
+EXPORT_SYMBOL(perf_pmu_name);
+
+int perf_num_counters(void)
+{
+       int num = 0;
+
+       if (cpum_cf_avail())
+               num += PERF_CPUM_CF_MAX_CTR;
+
+       return num;
+}
+EXPORT_SYMBOL(perf_num_counters);
+
+void perf_event_print_debug(void)
+{
+       struct cpumf_ctr_info cf_info;
+       unsigned long flags;
+       int cpu;
+
+       if (!cpum_cf_avail())
+               return;
+
+       local_irq_save(flags);
+
+       cpu = smp_processor_id();
+       memset(&cf_info, 0, sizeof(cf_info));
+       if (!qctri(&cf_info)) {
+               pr_info("CPU[%i] CPUM_CF: ver=%u.%u A=%04x E=%04x C=%04x\n",
+                       cpu, cf_info.cfvn, cf_info.csvn,
+                       cf_info.auth_ctl, cf_info.enable_ctl, cf_info.act_ctl);
+               print_hex_dump_bytes("CPUMF Query: ", DUMP_PREFIX_OFFSET,
+                                    &cf_info, sizeof(cf_info));
+       }
+
+       local_irq_restore(flags);
+}
+
+/* See also arch/s390/kernel/traps.c */
+static unsigned long __store_trace(struct perf_callchain_entry *entry,
+                                  unsigned long sp,
+                                  unsigned long low, unsigned long high)
+{
+       struct stack_frame *sf;
+       struct pt_regs *regs;
+
+       while (1) {
+               sp = sp & PSW_ADDR_INSN;
+               if (sp < low || sp > high - sizeof(*sf))
+                       return sp;
+               sf = (struct stack_frame *) sp;
+               perf_callchain_store(entry, sf->gprs[8] & PSW_ADDR_INSN);
+               /* Follow the backchain. */
+               while (1) {
+                       low = sp;
+                       sp = sf->back_chain & PSW_ADDR_INSN;
+                       if (!sp)
+                               break;
+                       if (sp <= low || sp > high - sizeof(*sf))
+                               return sp;
+                       sf = (struct stack_frame *) sp;
+                       perf_callchain_store(entry,
+                                            sf->gprs[8] & PSW_ADDR_INSN);
+               }
+               /* Zero backchain detected, check for interrupt frame. */
+               sp = (unsigned long) (sf + 1);
+               if (sp <= low || sp > high - sizeof(*regs))
+                       return sp;
+               regs = (struct pt_regs *) sp;
+               perf_callchain_store(entry, sf->gprs[8] & PSW_ADDR_INSN);
+               low = sp;
+               sp = regs->gprs[15];
+       }
+}
+
+void perf_callchain_kernel(struct perf_callchain_entry *entry,
+                          struct pt_regs *regs)
+{
+       unsigned long head;
+       struct stack_frame *head_sf;
+
+       if (user_mode(regs))
+               return;
+
+       head = regs->gprs[15];
+       head_sf = (struct stack_frame *) head;
+
+       if (!head_sf || !head_sf->back_chain)
+               return;
+
+       head = head_sf->back_chain;
+       head = __store_trace(entry, head, S390_lowcore.async_stack - ASYNC_SIZE,
+                            S390_lowcore.async_stack);
+
+       __store_trace(entry, head, S390_lowcore.thread_info,
+                     S390_lowcore.thread_info + THREAD_SIZE);
+}