+
+
+/* Return the maximum required space for all possible CPUs in case one
+ * CPU will be onlined during the START, READ, STOP cycles.
+ * To find out the size of the counter sets, any one CPU will do. They
+ * all have the same counter sets.
+ */
+static size_t cfset_needspace(unsigned int sets)
+{
+ struct cpu_cf_events *cpuhw = get_cpu_ptr(&cpu_cf_events);
+ size_t bytes = 0;
+ int i;
+
+ for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
+ if (!(sets & cpumf_ctr_ctl[i]))
+ continue;
+ bytes += cpum_cf_ctrset_size(i, &cpuhw->info) * sizeof(u64) +
+ sizeof(((struct s390_ctrset_setdata *)0)->set) +
+ sizeof(((struct s390_ctrset_setdata *)0)->no_cnts);
+ }
+ bytes = sizeof(((struct s390_ctrset_read *)0)->no_cpus) + nr_cpu_ids *
+ (bytes + sizeof(((struct s390_ctrset_cpudata *)0)->cpu_nr) +
+ sizeof(((struct s390_ctrset_cpudata *)0)->no_sets));
+ put_cpu_ptr(&cpu_cf_events);
+ return bytes;
+}
+
+static int cfset_all_copy(unsigned long arg, cpumask_t *mask)
+{
+ struct s390_ctrset_read __user *ctrset_read;
+ unsigned int cpu, cpus, rc;
+ void __user *uptr;
+
+ ctrset_read = (struct s390_ctrset_read __user *)arg;
+ uptr = ctrset_read->data;
+ for_each_cpu(cpu, mask) {
+ struct cpu_cf_events *cpuhw = per_cpu_ptr(&cpu_cf_events, cpu);
+ struct s390_ctrset_cpudata __user *ctrset_cpudata;
+
+ ctrset_cpudata = uptr;
+ rc = put_user(cpu, &ctrset_cpudata->cpu_nr);
+ rc |= put_user(cpuhw->sets, &ctrset_cpudata->no_sets);
+ rc |= copy_to_user(ctrset_cpudata->data, cpuhw->data,
+ cpuhw->used);
+ if (rc)
+ return -EFAULT;
+ uptr += sizeof(struct s390_ctrset_cpudata) + cpuhw->used;
+ cond_resched();
+ }
+ cpus = cpumask_weight(mask);
+ if (put_user(cpus, &ctrset_read->no_cpus))
+ return -EFAULT;
+ debug_sprintf_event(cf_dbg, 4, "%s copied %ld\n", __func__,
+ uptr - (void __user *)ctrset_read->data);
+ return 0;
+}
+
+static size_t cfset_cpuset_read(struct s390_ctrset_setdata *p, int ctrset,
+ int ctrset_size, size_t room)
+{
+ size_t need = 0;
+ int rc = -1;
+
+ need = sizeof(*p) + sizeof(u64) * ctrset_size;
+ if (need <= room) {
+ p->set = cpumf_ctr_ctl[ctrset];
+ p->no_cnts = ctrset_size;
+ rc = ctr_stcctm(ctrset, ctrset_size, (u64 *)p->cv);
+ if (rc == 3) /* Nothing stored */
+ need = 0;
+ }
+ return need;
+}
+
+/* Read all counter sets. */
+static void cfset_cpu_read(void *parm)
+{
+ struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
+ struct cfset_call_on_cpu_parm *p = parm;
+ int set, set_size;
+ size_t space;
+
+ /* No data saved yet */
+ cpuhw->used = 0;
+ cpuhw->sets = 0;
+ memset(cpuhw->data, 0, sizeof(cpuhw->data));
+
+ /* Scan the counter sets */
+ for (set = CPUMF_CTR_SET_BASIC; set < CPUMF_CTR_SET_MAX; ++set) {
+ struct s390_ctrset_setdata *sp = (void *)cpuhw->data +
+ cpuhw->used;
+
+ if (!(p->sets & cpumf_ctr_ctl[set]))
+ continue; /* Counter set not in list */
+ set_size = cpum_cf_ctrset_size(set, &cpuhw->info);
+ space = sizeof(cpuhw->data) - cpuhw->used;
+ space = cfset_cpuset_read(sp, set, set_size, space);
+ if (space) {
+ cpuhw->used += space;
+ cpuhw->sets += 1;
+ }
+ }
+ debug_sprintf_event(cf_dbg, 4, "%s sets %d used %zd\n", __func__,
+ cpuhw->sets, cpuhw->used);
+}
+
+static int cfset_all_read(unsigned long arg)
+{
+ struct cfset_call_on_cpu_parm p;
+ cpumask_var_t mask;
+ int rc;
+
+ if (!alloc_cpumask_var(&mask, GFP_KERNEL))
+ return -ENOMEM;
+
+ p.sets = cfset_request.ctrset;
+ cpumask_and(mask, &cfset_request.mask, cpu_online_mask);
+ on_each_cpu_mask(mask, cfset_cpu_read, &p, 1);
+ rc = cfset_all_copy(arg, mask);
+ free_cpumask_var(mask);
+ return rc;
+}
+
+static long cfset_ioctl_read(unsigned long arg)
+{
+ struct s390_ctrset_read read;
+ int ret = 0;
+
+ if (copy_from_user(&read, (char __user *)arg, sizeof(read)))
+ return -EFAULT;
+ ret = cfset_all_read(arg);
+ return ret;
+}
+
+static long cfset_ioctl_stop(void)
+{
+ int ret = ENXIO;
+
+ if (cfset_request.ctrset) {
+ ret = cfset_all_stop();
+ cfset_ctrset_clear();
+ }
+ return ret;
+}
+
+static long cfset_ioctl_start(unsigned long arg)
+{
+ struct s390_ctrset_start __user *ustart;
+ struct s390_ctrset_start start;
+ void __user *umask;
+ unsigned int len;
+ int ret = 0;
+ size_t need;
+
+ if (cfset_request.ctrset)
+ return -EBUSY;
+ ustart = (struct s390_ctrset_start __user *)arg;
+ if (copy_from_user(&start, ustart, sizeof(start)))
+ return -EFAULT;
+ if (start.version != S390_HWCTR_START_VERSION)
+ return -EINVAL;
+ if (start.counter_sets & ~(cpumf_ctr_ctl[CPUMF_CTR_SET_BASIC] |
+ cpumf_ctr_ctl[CPUMF_CTR_SET_USER] |
+ cpumf_ctr_ctl[CPUMF_CTR_SET_CRYPTO] |
+ cpumf_ctr_ctl[CPUMF_CTR_SET_EXT] |
+ cpumf_ctr_ctl[CPUMF_CTR_SET_MT_DIAG]))
+ return -EINVAL; /* Invalid counter set */
+ if (!start.counter_sets)
+ return -EINVAL; /* No counter set at all? */
+ cpumask_clear(&cfset_request.mask);
+ len = min_t(u64, start.cpumask_len, cpumask_size());
+ umask = (void __user *)start.cpumask;
+ if (copy_from_user(&cfset_request.mask, umask, len))
+ return -EFAULT;
+ if (cpumask_empty(&cfset_request.mask))
+ return -EINVAL;
+ need = cfset_needspace(start.counter_sets);
+ if (put_user(need, &ustart->data_bytes))
+ ret = -EFAULT;
+ if (ret)
+ goto out;
+ cfset_request.ctrset = start.counter_sets;
+ ret = cfset_all_start();
+out:
+ if (ret)
+ cfset_ctrset_clear();
+ debug_sprintf_event(cf_dbg, 4, "%s sets %#lx need %ld ret %d\n",
+ __func__, cfset_request.ctrset, need, ret);
+ return ret;
+}
+
+/* Entry point to the /dev/hwctr device interface.
+ * The ioctl system call supports three subcommands:
+ * S390_HWCTR_START: Start the specified counter sets on a CPU list. The
+ * counter set keeps running until explicitly stopped. Returns the number
+ * of bytes needed to store the counter values. If another S390_HWCTR_START
+ * ioctl subcommand is called without a previous S390_HWCTR_STOP stop
+ * command, -EBUSY is returned.
+ * S390_HWCTR_READ: Read the counter set values from specified CPU list given
+ * with the S390_HWCTR_START command.
+ * S390_HWCTR_STOP: Stops the counter sets on the CPU list given with the
+ * previous S390_HWCTR_START subcommand.
+ */
+static long cfset_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+{
+ int ret;
+
+ get_online_cpus();
+ mutex_lock(&cfset_ctrset_mutex);
+ switch (cmd) {
+ case S390_HWCTR_START:
+ ret = cfset_ioctl_start(arg);
+ break;
+ case S390_HWCTR_STOP:
+ ret = cfset_ioctl_stop();
+ break;
+ case S390_HWCTR_READ:
+ ret = cfset_ioctl_read(arg);
+ break;
+ default:
+ ret = -ENOTTY;
+ break;
+ }
+ mutex_unlock(&cfset_ctrset_mutex);
+ put_online_cpus();
+ return ret;
+}
+
+static const struct file_operations cfset_fops = {
+ .owner = THIS_MODULE,
+ .open = cfset_open,
+ .release = cfset_release,
+ .unlocked_ioctl = cfset_ioctl,
+ .compat_ioctl = cfset_ioctl,
+ .llseek = no_llseek
+};
+
+static struct miscdevice cfset_dev = {
+ .name = S390_HWCTR_DEVICE,
+ .minor = MISC_DYNAMIC_MINOR,
+ .fops = &cfset_fops,
+};
+
+int cfset_online_cpu(unsigned int cpu)
+{
+ struct cfset_call_on_cpu_parm p;
+
+ mutex_lock(&cfset_ctrset_mutex);
+ if (cfset_request.ctrset) {
+ p.sets = cfset_request.ctrset;
+ cfset_ioctl_on(&p);
+ cpumask_set_cpu(cpu, &cfset_request.mask);
+ }
+ mutex_unlock(&cfset_ctrset_mutex);
+ return 0;
+}
+
+int cfset_offline_cpu(unsigned int cpu)
+{
+ struct cfset_call_on_cpu_parm p;
+
+ mutex_lock(&cfset_ctrset_mutex);
+ if (cfset_request.ctrset) {
+ p.sets = cfset_request.ctrset;
+ cfset_ioctl_off(&p);
+ cpumask_clear_cpu(cpu, &cfset_request.mask);
+ }
+ mutex_unlock(&cfset_ctrset_mutex);
+ return 0;
+}
+
+static void cfdiag_read(struct perf_event *event)
+{
+ debug_sprintf_event(cf_dbg, 3, "%s event %#llx count %ld\n", __func__,
+ event->attr.config, local64_read(&event->count));
+}
+
+static int get_authctrsets(void)
+{
+ struct cpu_cf_events *cpuhw;
+ unsigned long auth = 0;
+ enum cpumf_ctr_set i;
+
+ cpuhw = &get_cpu_var(cpu_cf_events);
+ for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
+ if (cpuhw->info.auth_ctl & cpumf_ctr_ctl[i])
+ auth |= cpumf_ctr_ctl[i];
+ }
+ put_cpu_var(cpu_cf_events);
+ return auth;
+}
+
+/* Setup the event. Test for authorized counter sets and only include counter
+ * sets which are authorized at the time of the setup. Including unauthorized
+ * counter sets result in specification exception (and panic).
+ */
+static int cfdiag_event_init2(struct perf_event *event)
+{
+ struct perf_event_attr *attr = &event->attr;
+ int err = 0;
+
+ /* Set sample_period to indicate sampling */
+ event->hw.config = attr->config;
+ event->hw.sample_period = attr->sample_period;
+ local64_set(&event->hw.period_left, event->hw.sample_period);
+ local64_set(&event->count, 0);
+ event->hw.last_period = event->hw.sample_period;
+
+ /* Add all authorized counter sets to config_base. The
+ * the hardware init function is either called per-cpu or just once
+ * for all CPUS (event->cpu == -1). This depends on the whether
+ * counting is started for all CPUs or on a per workload base where
+ * the perf event moves from one CPU to another CPU.
+ * Checking the authorization on any CPU is fine as the hardware
+ * applies the same authorization settings to all CPUs.
+ */
+ event->hw.config_base = get_authctrsets();
+
+ /* No authorized counter sets, nothing to count/sample */
+ if (!event->hw.config_base)
+ err = -EINVAL;
+
+ debug_sprintf_event(cf_dbg, 5, "%s err %d config_base %#lx\n",
+ __func__, err, event->hw.config_base);
+ return err;
+}
+
+static int cfdiag_event_init(struct perf_event *event)
+{
+ struct perf_event_attr *attr = &event->attr;
+ int err = -ENOENT;
+
+ if (event->attr.config != PERF_EVENT_CPUM_CF_DIAG ||
+ event->attr.type != event->pmu->type)
+ goto out;
+
+ /* Raw events are used to access counters directly,
+ * hence do not permit excludes.
+ * This event is useless without PERF_SAMPLE_RAW to return counter set
+ * values as raw data.
+ */
+ if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv ||
+ !(attr->sample_type & (PERF_SAMPLE_CPU | PERF_SAMPLE_RAW))) {
+ err = -EOPNOTSUPP;
+ goto out;
+ }
+
+ /* Initialize for using the CPU-measurement counter facility */
+ cpumf_hw_inuse();
+ event->destroy = hw_perf_event_destroy;
+
+ err = cfdiag_event_init2(event);
+ if (unlikely(err))
+ event->destroy(event);
+out:
+ return err;
+}
+
+/* Create cf_diag/events/CF_DIAG event sysfs file. This counter is used
+ * to collect the complete counter sets for a scheduled process. Target
+ * are complete counter sets attached as raw data to the artificial event.
+ * This results in complete counter sets available when a process is
+ * scheduled. Contains the delta of every counter while the process was
+ * running.
+ */
+CPUMF_EVENT_ATTR(CF_DIAG, CF_DIAG, PERF_EVENT_CPUM_CF_DIAG);
+
+static struct attribute *cfdiag_events_attr[] = {
+ CPUMF_EVENT_PTR(CF_DIAG, CF_DIAG),
+ NULL,
+};
+
+PMU_FORMAT_ATTR(event, "config:0-63");
+
+static struct attribute *cfdiag_format_attr[] = {
+ &format_attr_event.attr,
+ NULL,
+};
+
+static struct attribute_group cfdiag_events_group = {
+ .name = "events",
+ .attrs = cfdiag_events_attr,
+};
+static struct attribute_group cfdiag_format_group = {
+ .name = "format",
+ .attrs = cfdiag_format_attr,
+};
+static const struct attribute_group *cfdiag_attr_groups[] = {
+ &cfdiag_events_group,
+ &cfdiag_format_group,
+ NULL,
+};
+
+/* Performance monitoring unit for event CF_DIAG. Since this event
+ * is also started and stopped via the perf_event_open() system call, use
+ * the same event enable/disable call back functions. They do not
+ * have a pointer to the perf_event strcture as first parameter.
+ *
+ * The functions XXX_add, XXX_del, XXX_start and XXX_stop are also common.
+ * Reuse them and distinguish the event (always first parameter) via
+ * 'config' member.
+ */
+static struct pmu cf_diag = {
+ .task_ctx_nr = perf_sw_context,
+ .event_init = cfdiag_event_init,
+ .pmu_enable = cpumf_pmu_enable,
+ .pmu_disable = cpumf_pmu_disable,
+ .add = cpumf_pmu_add,
+ .del = cpumf_pmu_del,
+ .start = cpumf_pmu_start,
+ .stop = cpumf_pmu_stop,
+ .read = cfdiag_read,
+
+ .attr_groups = cfdiag_attr_groups
+};
+
+/* Calculate memory needed to store all counter sets together with header and
+ * trailer data. This is independent of the counter set authorization which
+ * can vary depending on the configuration.
+ */
+static size_t cfdiag_maxsize(struct cpumf_ctr_info *info)
+{
+ size_t max_size = sizeof(struct cf_trailer_entry);
+ enum cpumf_ctr_set i;
+
+ for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
+ size_t size = cpum_cf_ctrset_size(i, info);
+
+ if (size)
+ max_size += size * sizeof(u64) +
+ sizeof(struct cf_ctrset_entry);
+ }
+ return max_size;
+}
+
+/* Get the CPU speed, try sampling facility first and CPU attributes second. */
+static void cfdiag_get_cpu_speed(void)
+{
+ if (cpum_sf_avail()) { /* Sampling facility first */
+ struct hws_qsi_info_block si;
+
+ memset(&si, 0, sizeof(si));
+ if (!qsi(&si)) {
+ cfdiag_cpu_speed = si.cpu_speed;
+ return;
+ }
+ }
+
+ /* Fallback: CPU speed extract static part. Used in case
+ * CPU Measurement Sampling Facility is turned off.
+ */
+ if (test_facility(34)) {
+ unsigned long mhz = __ecag(ECAG_CPU_ATTRIBUTE, 0);
+
+ if (mhz != -1UL)
+ cfdiag_cpu_speed = mhz & 0xffffffff;
+ }
+}
+
+static int cfset_init(void)
+{
+ struct cpumf_ctr_info info;
+ size_t need;
+ int rc;
+
+ if (qctri(&info))
+ return -ENODEV;
+
+ cfdiag_get_cpu_speed();
+ /* Make sure the counter set data fits into predefined buffer. */
+ need = cfdiag_maxsize(&info);
+ if (need > sizeof(((struct cpu_cf_events *)0)->start)) {
+ pr_err("Insufficient memory for PMU(cpum_cf_diag) need=%zu\n",
+ need);
+ return -ENOMEM;
+ }
+
+ rc = misc_register(&cfset_dev);
+ if (rc) {
+ pr_err("Registration of /dev/%s failed rc=%i\n",
+ cfset_dev.name, rc);
+ goto out;
+ }
+
+ rc = perf_pmu_register(&cf_diag, "cpum_cf_diag", -1);
+ if (rc) {
+ misc_deregister(&cfset_dev);
+ pr_err("Registration of PMU(cpum_cf_diag) failed with rc=%i\n",
+ rc);
+ }
+out:
+ return rc;
+}
+
+device_initcall(cpumf_pmu_init);