s390/perf: add support for the CPU-Measurement Sampling Facility
authorHendrik Brueckner <brueckner@linux.vnet.ibm.com>
Thu, 12 Dec 2013 15:32:47 +0000 (16:32 +0100)
committerMartin Schwidefsky <schwidefsky@de.ibm.com>
Mon, 16 Dec 2013 13:37:51 +0000 (14:37 +0100)
Introduce a perf PMU, "cpum_sf", to support the CPU-Measurement
Sampling Facility.  You can control the sampling facility through
this perf PMU interfaces.  Perf sampling events are created for
hardware samples.

For details about the CPU-Measurement Sampling Facility, see
"The Load-Program-Parameter and the CPU-Measurement Facilities" (SA23-2260).

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_sf.c [new file with mode: 0644]
arch/s390/kernel/perf_event.c

index f6dddeaad965d4ea174238e1a3ecd0acbc1ec297..d707abc26157de7ae6d11a642b8df40184ddec27 100644 (file)
@@ -210,6 +210,20 @@ static inline int lsctl(struct hws_lsctl_request_block *req)
 
 /* Sampling control helper functions */
 
+#include <linux/time.h>
+
+static inline unsigned long freq_to_sample_rate(struct hws_qsi_info_block *qsi,
+                                               unsigned long freq)
+{
+       return (USEC_PER_SEC / freq) * qsi->cpu_speed;
+}
+
+static inline unsigned long sample_rate_to_freq(struct hws_qsi_info_block *qsi,
+                                               unsigned long rate)
+{
+       return USEC_PER_SEC * qsi->cpu_speed / rate;
+}
+
 #define SDB_TE_ALERT_REQ_MASK  0x4000000000000000UL
 #define SDB_TE_BUFFER_FULL_MASK 0x8000000000000000UL
 
index 34185020ae0aaa4defcb9254ba0afebac20074ef..b4eea25f379efa53040542819bd66e2b2d466047 100644 (file)
 #include <linux/device.h>
 #include <asm/cpu_mf.h>
 
-/* CPU-measurement counter facility */
-#define PERF_CPUM_CF_MAX_CTR           256
-
 /* Per-CPU flags for PMU states */
 #define PMU_F_RESERVED                 0x1000
 #define PMU_F_ENABLED                  0x2000
+#define PMU_F_IN_USE                   0x4000
+#define PMU_F_ERR_IBE                  0x0100
+#define PMU_F_ERR_LSDA                 0x0200
+#define PMU_F_ERR_MASK                 (PMU_F_ERR_IBE|PMU_F_ERR_LSDA)
 
 /* Perf defintions for PMU event attributes in sysfs */
 extern __init const struct attribute_group **cpumf_cf_event_group(void);
@@ -41,5 +42,15 @@ extern unsigned long perf_instruction_pointer(struct pt_regs *regs);
 extern unsigned long perf_misc_flags(struct pt_regs *regs);
 #define perf_misc_flags(regs) perf_misc_flags(regs)
 
+/* Perf PMU definitions for the counter facility */
+#define PERF_CPUM_CF_MAX_CTR           256
+
+/* Perf PMU definitions for the sampling facility */
+#define PERF_CPUM_SF_MAX_CTR           1
+#define PERF_EVENT_CPUM_SF             0xB0000UL       /* Raw event ID */
+
+#define TEAR_REG(hwc)          ((hwc)->last_tag)
+#define SAMPL_RATE(hwc)                ((hwc)->event_base)
+
 #endif /* CONFIG_64BIT */
 #endif /* _ASM_S390_PERF_EVENT_H */
index 9f1e2adbd77e4b73431e5e4299267b83a7f1c603..1b3ac09c11b6df2981ca3dbde3cc6df3fba5d2e9 100644 (file)
@@ -60,7 +60,7 @@ obj-$(CONFIG_FTRACE_SYSCALLS)  += ftrace.o
 obj-$(CONFIG_CRASH_DUMP)       += crash_dump.o
 
 ifdef CONFIG_64BIT
-obj-$(CONFIG_PERF_EVENTS)      += perf_event.o perf_cpum_cf.o \
+obj-$(CONFIG_PERF_EVENTS)      += perf_event.o perf_cpum_cf.o perf_cpum_sf.o \
                                                perf_cpum_cf_events.o
 obj-y                          += runtime_instr.o cache.o
 endif
diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c
new file mode 100644 (file)
index 0000000..141eca0
--- /dev/null
@@ -0,0 +1,1024 @@
+/*
+ * Performance event support for the System z CPU-measurement Sampling Facility
+ *
+ * Copyright IBM Corp. 2013
+ * 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_sf"
+#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/export.h>
+#include <asm/cpu_mf.h>
+#include <asm/irq.h>
+#include <asm/debug.h>
+#include <asm/timex.h>
+
+/* Minimum number of sample-data-block-tables:
+ * At least one table is required for the sampling buffer structure.
+ * A single table contains up to 511 pointers to sample-data-blocks.
+ */
+#define CPUM_SF_MIN_SDBT    1
+
+/* Minimum number of sample-data-blocks:
+ * The minimum designates a single page for sample-data-block, i.e.,
+ * up to 126 sample-data-blocks with a size of 32 bytes (bsdes).
+ */
+#define CPUM_SF_MIN_SDB            126
+
+/* Maximum number of sample-data-blocks:
+ * The maximum number designates approx. 256K per CPU including
+ * the given number of sample-data-blocks and taking the number
+ * of sample-data-block tables into account.
+ *
+ * Later, this number can be increased for extending the sampling
+ * buffer, for example, by factor 2 (512K) or 4 (1M).
+ */
+#define CPUM_SF_MAX_SDB            6471
+
+struct sf_buffer {
+       unsigned long    sdbt;      /* Sample-data-block-table origin */
+       /* buffer characteristics (required for buffer increments) */
+       unsigned long num_sdb;      /* Number of sample-data-blocks */
+       unsigned long    tail;      /* last sample-data-block-table */
+};
+
+struct cpu_hw_sf {
+       /* CPU-measurement sampling information block */
+       struct hws_qsi_info_block qsi;
+       struct hws_lsctl_request_block lsctl;
+       struct sf_buffer sfb;       /* Sampling buffer */
+       unsigned int flags;         /* Status flags */
+       struct perf_event *event;   /* Scheduled perf event */
+};
+static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf);
+
+/* Debug feature */
+static debug_info_t *sfdbg;
+
+/*
+ * sf_buffer_available() - Check for an allocated sampling buffer
+ */
+static int sf_buffer_available(struct cpu_hw_sf *cpuhw)
+{
+       return (cpuhw->sfb.sdbt) ? 1 : 0;
+}
+
+/*
+ * deallocate sampling facility buffer
+ */
+static void free_sampling_buffer(struct sf_buffer *sfb)
+{
+       unsigned long sdbt, *curr;
+
+       if (!sfb->sdbt)
+               return;
+
+       sdbt = sfb->sdbt;
+       curr = (unsigned long *) sdbt;
+
+       /* we'll free the SDBT after all SDBs are processed... */
+       while (1) {
+               if (!*curr || !sdbt)
+                       break;
+
+               /* watch for link entry reset if found */
+               if (is_link_entry(curr)) {
+                       curr = get_next_sdbt(curr);
+                       if (sdbt)
+                               free_page(sdbt);
+
+                       /* we are done if we reach the origin */
+                       if ((unsigned long) curr == sfb->sdbt)
+                               break;
+                       else
+                               sdbt = (unsigned long) curr;
+               } else {
+                       /* process SDB pointer */
+                       if (*curr) {
+                               free_page(*curr);
+                               curr++;
+                       }
+               }
+       }
+
+       debug_sprintf_event(sfdbg, 5,
+                           "free_sampling_buffer: freed sdbt=%0lx\n", sfb->sdbt);
+       memset(sfb, 0, sizeof(*sfb));
+}
+
+/*
+ * allocate_sampling_buffer() - allocate sampler memory
+ *
+ * Allocates and initializes a sampling buffer structure using the
+ * specified number of sample-data-blocks (SDB).  For each allocation,
+ * a 4K page is used.  The number of sample-data-block-tables (SDBT)
+ * are calculated from SDBs.
+ * Also set the ALERT_REQ mask in each SDBs trailer.
+ *
+ * Returns zero on success, non-zero otherwise.
+ */
+static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
+{
+       int j, k, rc;
+       unsigned long *sdbt, *tail, *trailer;
+       unsigned long sdb;
+       unsigned long num_sdbt, sdb_per_table;
+
+       if (sfb->sdbt)
+               return -EINVAL;
+       sfb->num_sdb = 0;
+
+       /* Compute the number of required sample-data-block-tables (SDBT) */
+       num_sdbt = num_sdb / ((PAGE_SIZE - 8) / 8);
+       if (num_sdbt < CPUM_SF_MIN_SDBT)
+               num_sdbt = CPUM_SF_MIN_SDBT;
+       sdb_per_table = (PAGE_SIZE - 8) / 8;
+
+       debug_sprintf_event(sfdbg, 4, "alloc_sampling_buffer: num_sdbt=%lu "
+                           "num_sdb=%lu sdb_per_table=%lu\n",
+                           num_sdbt, num_sdb, sdb_per_table);
+       sdbt = NULL;
+       tail = sdbt;
+
+       for (j = 0; j < num_sdbt; j++) {
+               sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
+               if (!sdbt) {
+                       rc = -ENOMEM;
+                       goto allocate_sdbt_error;
+               }
+
+               /* save origin of sample-data-block-table */
+               if (!sfb->sdbt)
+                       sfb->sdbt = (unsigned long) sdbt;
+
+               /* link current page to tail of chain */
+               if (tail)
+                       *tail = (unsigned long)(void *) sdbt + 1;
+
+               for (k = 0; k < num_sdb && k < sdb_per_table; k++) {
+                       /* get and set SDB page */
+                       sdb = get_zeroed_page(GFP_KERNEL);
+                       if (!sdb) {
+                               rc = -ENOMEM;
+                               goto allocate_sdbt_error;
+                       }
+                       *sdbt = sdb;
+                       trailer = trailer_entry_ptr(*sdbt);
+                       *trailer = SDB_TE_ALERT_REQ_MASK;
+                       sdbt++;
+               }
+               num_sdb -= k;
+               sfb->num_sdb += k;      /* count allocated sdb's */
+               tail = sdbt;
+       }
+
+       rc = 0;
+       if (tail)
+               *tail = sfb->sdbt + 1;
+       sfb->tail = (unsigned long) (void *)tail;
+
+allocate_sdbt_error:
+       if (rc)
+               free_sampling_buffer(sfb);
+       else
+               debug_sprintf_event(sfdbg, 4,
+                       "alloc_sampling_buffer: tear=%0lx dear=%0lx\n",
+                       sfb->sdbt, *(unsigned long *) sfb->sdbt);
+       return rc;
+}
+
+static int allocate_sdbt(struct cpu_hw_sf *cpuhw, const struct hw_perf_event *hwc)
+{
+       unsigned long n_sdb, freq;
+       unsigned long factor;
+
+       /* Calculate sampling buffers using 4K pages
+        *
+        *    1. Use frequency as input.  The samping buffer is designed for
+        *       a complete second.  This can be adjusted through the "factor"
+        *       variable.
+        *       In any case, alloc_sampling_buffer() sets the Alert Request
+        *       Control indicator to trigger measurement-alert to harvest
+        *       sample-data-blocks (sdb).
+        *
+        *    2. Compute the number of sample-data-blocks and ensure a minimum
+        *       of CPUM_SF_MIN_SDB.  Also ensure the upper limit does not
+        *       exceed CPUM_SF_MAX_SDB.  See also the remarks for these
+        *       symbolic constants.
+        *
+        *    3. Compute number of pages used for the sample-data-block-table
+        *       and ensure a minimum of CPUM_SF_MIN_SDBT (at minimum one table
+        *       to manage up to 511 sample-data-blocks).
+        */
+       freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
+       factor = 1;
+       n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / cpuhw->qsi.bsdes));
+       if (n_sdb < CPUM_SF_MIN_SDB)
+               n_sdb = CPUM_SF_MIN_SDB;
+
+       /* Return if there is already a sampling buffer allocated.
+        * XXX Remove this later and check number of available and
+        * required sdb's and, if necessary, increase the sampling buffer.
+        */
+       if (sf_buffer_available(cpuhw))
+               return 0;
+
+       debug_sprintf_event(sfdbg, 3,
+                           "allocate_sdbt: rate=%lu f=%lu sdb=%lu/%i cpuhw=%p\n",
+                           SAMPL_RATE(hwc), freq, n_sdb, CPUM_SF_MAX_SDB, cpuhw);
+
+       return alloc_sampling_buffer(&cpuhw->sfb,
+                              min_t(unsigned long, n_sdb, CPUM_SF_MAX_SDB));
+}
+
+
+/* Number of perf events counting hardware events */
+static atomic_t num_events;
+/* Used to avoid races in calling reserve/release_cpumf_hardware */
+static DEFINE_MUTEX(pmc_reserve_mutex);
+
+/*
+ * sf_disable() - Switch off sampling facility
+ */
+static int sf_disable(void)
+{
+       struct hws_lsctl_request_block sreq;
+
+       memset(&sreq, 0, sizeof(sreq));
+       return lsctl(&sreq);
+}
+
+
+#define PMC_INIT      0
+#define PMC_RELEASE   1
+static void setup_pmc_cpu(void *flags)
+{
+       int err;
+       struct cpu_hw_sf *cpusf = &__get_cpu_var(cpu_hw_sf);
+
+       /* XXX Improve error handling and pass a flag in the *flags
+        *     variable to indicate failures.  Alternatively, ignore
+        *     (print) errors here and let the PMU functions fail if
+        *     the per-cpu PMU_F_RESERVED flag is not.
+        */
+       err = 0;
+       switch (*((int *) flags)) {
+       case PMC_INIT:
+               memset(cpusf, 0, sizeof(*cpusf));
+               err = qsi(&cpusf->qsi);
+               if (err)
+                       break;
+               cpusf->flags |= PMU_F_RESERVED;
+               err = sf_disable();
+               if (err)
+                       pr_err("Switching off the sampling facility failed "
+                              "with rc=%i\n", err);
+               debug_sprintf_event(sfdbg, 5,
+                                   "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf);
+               break;
+       case PMC_RELEASE:
+               cpusf->flags &= ~PMU_F_RESERVED;
+               err = sf_disable();
+               if (err) {
+                       pr_err("Switching off the sampling facility failed "
+                              "with rc=%i\n", err);
+               } else {
+                       if (cpusf->sfb.sdbt)
+                               free_sampling_buffer(&cpusf->sfb);
+               }
+               debug_sprintf_event(sfdbg, 5,
+                                   "setup_pmc_cpu: released: cpuhw=%p\n", cpusf);
+               break;
+       }
+}
+
+static void release_pmc_hardware(void)
+{
+       int flags = PMC_RELEASE;
+
+       irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
+       on_each_cpu(setup_pmc_cpu, &flags, 1);
+}
+
+static int reserve_pmc_hardware(void)
+{
+       int flags = PMC_INIT;
+
+       on_each_cpu(setup_pmc_cpu, &flags, 1);
+       irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
+
+       return 0;
+}
+
+static void hw_perf_event_destroy(struct perf_event *event)
+{
+       /* Release PMC if this is the last perf 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);
+       }
+}
+
+static void hw_init_period(struct hw_perf_event *hwc, u64 period)
+{
+       hwc->sample_period = period;
+       hwc->last_period = hwc->sample_period;
+       local64_set(&hwc->period_left, hwc->sample_period);
+}
+
+static void hw_reset_registers(struct hw_perf_event *hwc,
+                              unsigned long sdbt_origin)
+{
+       TEAR_REG(hwc) = sdbt_origin;          /* (re)set to first sdb table */
+}
+
+static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
+                                  unsigned long rate)
+{
+       if (rate < si->min_sampl_rate)
+               return si->min_sampl_rate;
+       if (rate > si->max_sampl_rate)
+               return si->max_sampl_rate;
+       return rate;
+}
+
+static int __hw_perf_event_init(struct perf_event *event)
+{
+       struct cpu_hw_sf *cpuhw;
+       struct hws_qsi_info_block si;
+       struct perf_event_attr *attr = &event->attr;
+       struct hw_perf_event *hwc = &event->hw;
+       unsigned long rate;
+       int cpu, err;
+
+       /* Reserve CPU-measurement sampling facility */
+       err = 0;
+       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;
+
+       if (err)
+               goto out;
+
+       /* Access per-CPU sampling information (query sampling info) */
+       /*
+        * The event->cpu value can be -1 to count on every CPU, for example,
+        * when attaching to a task.  If this is specified, use the query
+        * sampling info from the current CPU, otherwise use event->cpu to
+        * retrieve the per-CPU information.
+        * Later, cpuhw indicates whether to allocate sampling buffers for a
+        * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
+        */
+       memset(&si, 0, sizeof(si));
+       cpuhw = NULL;
+       if (event->cpu == -1)
+               qsi(&si);
+       else {
+               /* Event is pinned to a particular CPU, retrieve the per-CPU
+                * sampling structure for accessing the CPU-specific QSI.
+                */
+               cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
+               si = cpuhw->qsi;
+       }
+
+       /* Check sampling facility authorization and, if not authorized,
+        * fall back to other PMUs.  It is safe to check any CPU because
+        * the authorization is identical for all configured CPUs.
+        */
+       if (!si.as) {
+               err = -ENOENT;
+               goto out;
+       }
+
+       /* The sampling information (si) contains information about the
+        * min/max sampling intervals and the CPU speed.  So calculate the
+        * correct sampling interval and avoid the whole period adjust
+        * feedback loop.
+        */
+       rate = 0;
+       if (attr->freq) {
+               rate = freq_to_sample_rate(&si, attr->sample_freq);
+               rate = hw_limit_rate(&si, rate);
+               attr->freq = 0;
+               attr->sample_period = rate;
+       } else {
+               /* The min/max sampling rates specifies the valid range
+                * of sample periods.  If the specified sample period is
+                * out of range, limit the period to the range boundary.
+                */
+               rate = hw_limit_rate(&si, hwc->sample_period);
+
+               /* The perf core maintains a maximum sample rate that is
+                * configurable through the sysctl interface.  Ensure the
+                * sampling rate does not exceed this value.  This also helps
+                * to avoid throttling when pushing samples with
+                * perf_event_overflow().
+                */
+               if (sample_rate_to_freq(&si, rate) >
+                     sysctl_perf_event_sample_rate) {
+                       err = -EINVAL;
+                       debug_sprintf_event(sfdbg, 1, "Sampling rate exceeds maximum perf sample rate\n");
+                       goto out;
+               }
+       }
+       SAMPL_RATE(hwc) = rate;
+       hw_init_period(hwc, SAMPL_RATE(hwc));
+
+       /* Allocate the per-CPU sampling buffer using the CPU information
+        * from the event.  If the event is not pinned to a particular
+        * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
+        * buffers for each online CPU.
+        */
+       if (cpuhw)
+               /* Event is pinned to a particular CPU */
+               err = allocate_sdbt(cpuhw, hwc);
+       else {
+               /* Event is not pinned, allocate sampling buffer on
+                * each online CPU
+                */
+               for_each_online_cpu(cpu) {
+                       cpuhw = &per_cpu(cpu_hw_sf, cpu);
+                       err = allocate_sdbt(cpuhw, hwc);
+                       if (err)
+                               break;
+               }
+       }
+out:
+       return err;
+}
+
+static int cpumsf_pmu_event_init(struct perf_event *event)
+{
+       int err;
+
+       if (event->attr.type != PERF_TYPE_RAW)
+               return -ENOENT;
+
+       if (event->attr.config != PERF_EVENT_CPUM_SF)
+               return -ENOENT;
+
+       if (event->cpu >= nr_cpumask_bits ||
+           (event->cpu >= 0 && !cpu_online(event->cpu)))
+               return -ENODEV;
+
+       err = __hw_perf_event_init(event);
+       if (unlikely(err))
+               if (event->destroy)
+                       event->destroy(event);
+       return err;
+}
+
+static void cpumsf_pmu_enable(struct pmu *pmu)
+{
+       struct cpu_hw_sf *cpuhw = &__get_cpu_var(cpu_hw_sf);
+       int err;
+
+       if (cpuhw->flags & PMU_F_ENABLED)
+               return;
+
+       if (cpuhw->flags & PMU_F_ERR_MASK)
+               return;
+
+       cpuhw->flags |= PMU_F_ENABLED;
+       barrier();
+
+       err = lsctl(&cpuhw->lsctl);
+       if (err) {
+               cpuhw->flags &= ~PMU_F_ENABLED;
+               pr_err("Loading sampling controls failed: op=%i err=%i\n",
+                       1, err);
+               return;
+       }
+
+       debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i tear=%p dear=%p\n",
+                           cpuhw->lsctl.es, cpuhw->lsctl.cs,
+                           (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
+}
+
+static void cpumsf_pmu_disable(struct pmu *pmu)
+{
+       struct cpu_hw_sf *cpuhw = &__get_cpu_var(cpu_hw_sf);
+       struct hws_lsctl_request_block inactive;
+       struct hws_qsi_info_block si;
+       int err;
+
+       if (!(cpuhw->flags & PMU_F_ENABLED))
+               return;
+
+       if (cpuhw->flags & PMU_F_ERR_MASK)
+               return;
+
+       /* Switch off sampling activation control */
+       inactive = cpuhw->lsctl;
+       inactive.cs = 0;
+
+       err = lsctl(&inactive);
+       if (err) {
+               pr_err("Loading sampling controls failed: op=%i err=%i\n",
+                       2, err);
+               return;
+       }
+
+       /* Save state of TEAR and DEAR register contents */
+       if (!qsi(&si)) {
+               /* TEAR/DEAR values are valid only if the sampling facility is
+                * enabled.  Note that cpumsf_pmu_disable() might be called even
+                * for a disabled sampling facility because cpumsf_pmu_enable()
+                * controls the enable/disable state.
+                */
+               if (si.es) {
+                       cpuhw->lsctl.tear = si.tear;
+                       cpuhw->lsctl.dear = si.dear;
+               }
+       } else
+               debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
+                                   "qsi() failed with err=%i\n", err);
+
+       cpuhw->flags &= ~PMU_F_ENABLED;
+}
+
+/* perf_push_sample() - Push samples to perf
+ * @event:     The perf event
+ * @sample:    Hardware sample data
+ *
+ * Use the hardware sample data to create perf event sample.  The sample
+ * is the pushed to the event subsystem and the function checks for
+ * possible event overflows.  If an event overflow occurs, the PMU is
+ * stopped.
+ *
+ * Return non-zero if an event overflow occurred.
+ */
+static int perf_push_sample(struct perf_event *event,
+                           struct hws_data_entry *sample)
+{
+       int overflow;
+       struct pt_regs regs;
+       struct perf_sample_data data;
+
+       /* Skip samples that are invalid or for which the instruction address
+        * is not predictable.  For the latter, the wait-state bit is set.
+        */
+       if (sample->I || sample->W)
+               return 0;
+
+       perf_sample_data_init(&data, 0, event->hw.last_period);
+
+       memset(&regs, 0, sizeof(regs));
+       regs.psw.addr = sample->ia;
+       if (sample->T)
+               regs.psw.mask |= PSW_MASK_DAT;
+       if (sample->W)
+               regs.psw.mask |= PSW_MASK_WAIT;
+       if (sample->P)
+               regs.psw.mask |= PSW_MASK_PSTATE;
+       switch (sample->AS) {
+       case 0x0:
+               regs.psw.mask |= PSW_ASC_PRIMARY;
+               break;
+       case 0x1:
+               regs.psw.mask |= PSW_ASC_ACCREG;
+               break;
+       case 0x2:
+               regs.psw.mask |= PSW_ASC_SECONDARY;
+               break;
+       case 0x3:
+               regs.psw.mask |= PSW_ASC_HOME;
+               break;
+       }
+
+       overflow = 0;
+       if (perf_event_overflow(event, &data, &regs)) {
+               overflow = 1;
+               event->pmu->stop(event, 0);
+               debug_sprintf_event(sfdbg, 4, "perf_push_sample: PMU stopped"
+                                   " because of an event overflow\n");
+       }
+       perf_event_update_userpage(event);
+
+       return overflow;
+}
+
+static void perf_event_count_update(struct perf_event *event, u64 count)
+{
+       local64_add(count, &event->count);
+}
+
+/* hw_collect_samples() - Walk through a sample-data-block and collect samples
+ * @event:     The perf event
+ * @sdbt:      Sample-data-block table
+ * @overflow:  Event overflow counter
+ *
+ * Walks through a sample-data-block and collects hardware sample-data that is
+ * pushed to the perf event subsystem. The overflow reports the number of
+ * samples that has been discarded due to an event overflow.
+ */
+static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
+                              unsigned long long *overflow)
+{
+       struct hws_data_entry *sample;
+       unsigned long *trailer;
+
+       trailer = trailer_entry_ptr(*sdbt);
+       sample = (struct hws_data_entry *) *sdbt;
+       while ((unsigned long *) sample < trailer) {
+               /* Check for an empty sample */
+               if (!sample->def)
+                       break;
+
+               /* Update perf event period */
+               perf_event_count_update(event, SAMPL_RATE(&event->hw));
+
+               /* Check for basic sampling mode */
+               if (sample->def == 0x0001) {
+                       /* If an event overflow occurred, the PMU is stopped to
+                        * throttle event delivery.  Remaining sample data is
+                        * discarded.
+                        */
+                       if (!*overflow)
+                               *overflow = perf_push_sample(event, sample);
+                       else
+                               /* Count discarded samples */
+                               *overflow += 1;
+               } else
+                       /* Sample slot is not yet written or other record */
+                       debug_sprintf_event(sfdbg, 5, "hw_collect_samples: "
+                                           "Unknown sample data entry format:"
+                                           " %i\n", sample->def);
+
+               /* Reset sample slot and advance to next sample */
+               sample->def = 0;
+               sample++;
+       }
+}
+
+/* hw_perf_event_update() - Process sampling buffer
+ * @event:     The perf event
+ * @flush_all: Flag to also flush partially filled sample-data-blocks
+ *
+ * Processes the sampling buffer and create perf event samples.
+ * The sampling buffer position are retrieved and saved in the TEAR_REG
+ * register of the specified perf event.
+ *
+ * Only full sample-data-blocks are processed. Specify the flash_all flag
+ * to also walk through partially filled sample-data-blocks.
+ *
+ */
+static void hw_perf_event_update(struct perf_event *event, int flush_all)
+{
+       struct hw_perf_event *hwc = &event->hw;
+       struct hws_trailer_entry *te;
+       unsigned long *sdbt;
+       unsigned long long event_overflow, sampl_overflow;
+       int done;
+
+       sdbt = (unsigned long *) TEAR_REG(hwc);
+       done = event_overflow = sampl_overflow = 0;
+       while (!done) {
+               /* Get the trailer entry of the sample-data-block */
+               te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
+
+               /* Leave loop if no more work to do (block full indicator) */
+               if (!te->f) {
+                       done = 1;
+                       if (!flush_all)
+                               break;
+               }
+
+               /* Check sample overflow count */
+               if (te->overflow) {
+                       /* Increment sample overflow counter */
+                       sampl_overflow += te->overflow;
+
+                       /* XXX: If an sample overflow occurs, increase the
+                        *      sampling buffer.  Set a "realloc" flag because
+                        *      the sampler must be re-enabled for changing
+                        *      the sample-data-block-table content.
+                        */
+               }
+
+               /* Timestamps are valid for full sample-data-blocks only */
+               debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
+                                   "overflow=%llu timestamp=0x%llx\n",
+                                   sdbt, te->overflow,
+                                   (te->f) ? te->timestamp : 0ULL);
+
+               /* Collect all samples from a single sample-data-block and
+                * flag if an (perf) event overflow happened.  If so, the PMU
+                * is stopped and remaining samples will be discarded.
+                */
+               hw_collect_samples(event, sdbt, &event_overflow);
+
+               /* Reset trailer */
+               xchg(&te->overflow, 0);
+               xchg((unsigned char *) te, 0x40);
+
+               /* Advance to next sample-data-block */
+               sdbt++;
+               if (is_link_entry(sdbt))
+                       sdbt = get_next_sdbt(sdbt);
+
+               /* Update event hardware registers */
+               TEAR_REG(hwc) = (unsigned long) sdbt;
+
+               /* Stop processing sample-data if all samples of the current
+                * sample-data-block were flushed even if it was not full.
+                */
+               if (flush_all && done)
+                       break;
+
+               /* If an event overflow happened, discard samples by
+                * processing any remaining sample-data-blocks.
+                */
+               if (event_overflow)
+                       flush_all = 1;
+       }
+
+       if (sampl_overflow || event_overflow)
+               debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
+                                   "overflow stats: sample=%llu event=%llu\n",
+                                   sampl_overflow, event_overflow);
+}
+
+static void cpumsf_pmu_read(struct perf_event *event)
+{
+       /* Nothing to do ... updates are interrupt-driven */
+}
+
+/* Activate sampling control.
+ * Next call of pmu_enable() starts sampling.
+ */
+static void cpumsf_pmu_start(struct perf_event *event, int flags)
+{
+       struct cpu_hw_sf *cpuhw = &__get_cpu_var(cpu_hw_sf);
+
+       if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
+               return;
+
+       if (flags & PERF_EF_RELOAD)
+               WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
+
+       perf_pmu_disable(event->pmu);
+       event->hw.state = 0;
+       cpuhw->lsctl.cs = 1;
+       perf_pmu_enable(event->pmu);
+}
+
+/* Deactivate sampling control.
+ * Next call of pmu_enable() stops sampling.
+ */
+static void cpumsf_pmu_stop(struct perf_event *event, int flags)
+{
+       struct cpu_hw_sf *cpuhw = &__get_cpu_var(cpu_hw_sf);
+
+       if (event->hw.state & PERF_HES_STOPPED)
+               return;
+
+       perf_pmu_disable(event->pmu);
+       cpuhw->lsctl.cs = 0;
+       event->hw.state |= PERF_HES_STOPPED;
+
+       if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
+               hw_perf_event_update(event, 1);
+               event->hw.state |= PERF_HES_UPTODATE;
+       }
+       perf_pmu_enable(event->pmu);
+}
+
+static int cpumsf_pmu_add(struct perf_event *event, int flags)
+{
+       struct cpu_hw_sf *cpuhw = &__get_cpu_var(cpu_hw_sf);
+       int err;
+
+       if (cpuhw->flags & PMU_F_IN_USE)
+               return -EAGAIN;
+
+       if (!cpuhw->sfb.sdbt)
+               return -EINVAL;
+
+       err = 0;
+       perf_pmu_disable(event->pmu);
+
+       event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
+
+       /* Set up sampling controls.  Always program the sampling register
+        * using the SDB-table start.  Reset TEAR_REG event hardware register
+        * that is used by hw_perf_event_update() to store the sampling buffer
+        * position after samples have been flushed.
+        */
+       cpuhw->lsctl.s = 0;
+       cpuhw->lsctl.h = 1;
+       cpuhw->lsctl.tear = cpuhw->sfb.sdbt;
+       cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
+       cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
+       hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);
+
+       /* Ensure sampling functions are in the disabled state.  If disabled,
+        * switch on sampling enable control. */
+       if (WARN_ON_ONCE(cpuhw->lsctl.es == 1)) {
+               err = -EAGAIN;
+               goto out;
+       }
+       cpuhw->lsctl.es = 1;
+
+       /* Set in_use flag and store event */
+       event->hw.idx = 0;        /* only one sampling event per CPU supported */
+       cpuhw->event = event;
+       cpuhw->flags |= PMU_F_IN_USE;
+
+       if (flags & PERF_EF_START)
+               cpumsf_pmu_start(event, PERF_EF_RELOAD);
+out:
+       perf_event_update_userpage(event);
+       perf_pmu_enable(event->pmu);
+       return err;
+}
+
+static void cpumsf_pmu_del(struct perf_event *event, int flags)
+{
+       struct cpu_hw_sf *cpuhw = &__get_cpu_var(cpu_hw_sf);
+
+       perf_pmu_disable(event->pmu);
+       cpumsf_pmu_stop(event, PERF_EF_UPDATE);
+
+       cpuhw->lsctl.es = 0;
+       cpuhw->flags &= ~PMU_F_IN_USE;
+       cpuhw->event = NULL;
+
+       perf_event_update_userpage(event);
+       perf_pmu_enable(event->pmu);
+}
+
+static int cpumsf_pmu_event_idx(struct perf_event *event)
+{
+       return event->hw.idx;
+}
+
+CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
+
+static struct attribute *cpumsf_pmu_events_attr[] = {
+       CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC),
+       NULL,
+};
+
+PMU_FORMAT_ATTR(event, "config:0-63");
+
+static struct attribute *cpumsf_pmu_format_attr[] = {
+       &format_attr_event.attr,
+       NULL,
+};
+
+static struct attribute_group cpumsf_pmu_events_group = {
+       .name = "events",
+       .attrs = cpumsf_pmu_events_attr,
+};
+static struct attribute_group cpumsf_pmu_format_group = {
+       .name = "format",
+       .attrs = cpumsf_pmu_format_attr,
+};
+static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
+       &cpumsf_pmu_events_group,
+       &cpumsf_pmu_format_group,
+       NULL,
+};
+
+static struct pmu cpumf_sampling = {
+       .pmu_enable   = cpumsf_pmu_enable,
+       .pmu_disable  = cpumsf_pmu_disable,
+
+       .event_init   = cpumsf_pmu_event_init,
+       .add          = cpumsf_pmu_add,
+       .del          = cpumsf_pmu_del,
+
+       .start        = cpumsf_pmu_start,
+       .stop         = cpumsf_pmu_stop,
+       .read         = cpumsf_pmu_read,
+
+       .event_idx    = cpumsf_pmu_event_idx,
+       .attr_groups  = cpumsf_pmu_attr_groups,
+};
+
+static void cpumf_measurement_alert(struct ext_code ext_code,
+                                   unsigned int alert, unsigned long unused)
+{
+       struct cpu_hw_sf *cpuhw;
+
+       if (!(alert & CPU_MF_INT_SF_MASK))
+               return;
+       inc_irq_stat(IRQEXT_CMS);
+       cpuhw = &__get_cpu_var(cpu_hw_sf);
+
+       /* 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;
+
+       /* The processing below must take care of multiple alert events that
+        * might be indicated concurrently. */
+
+       /* Program alert request */
+       if (alert & CPU_MF_INT_SF_PRA) {
+               if (cpuhw->flags & PMU_F_IN_USE)
+                       hw_perf_event_update(cpuhw->event, 0);
+               else
+                       WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
+       }
+
+       /* Report measurement alerts only for non-PRA codes */
+       if (alert != CPU_MF_INT_SF_PRA)
+               debug_sprintf_event(sfdbg, 6, "measurement alert: 0x%x\n", alert);
+
+       /* Sampling authorization change request */
+       if (alert & CPU_MF_INT_SF_SACA)
+               qsi(&cpuhw->qsi);
+
+       /* Loss of sample data due to high-priority machine activities */
+       if (alert & CPU_MF_INT_SF_LSDA) {
+               pr_err("Sample data was lost\n");
+               cpuhw->flags |= PMU_F_ERR_LSDA;
+               sf_disable();
+       }
+
+       /* Invalid sampling buffer entry */
+       if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
+               pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
+                      alert);
+               cpuhw->flags |= PMU_F_ERR_IBE;
+               sf_disable();
+       }
+}
+
+static int __cpuinit cpumf_pmu_notifier(struct notifier_block *self,
+                                       unsigned long action, void *hcpu)
+{
+       unsigned int cpu = (long) hcpu;
+       int flags;
+
+       /* Ignore the notification if no events are scheduled on the PMU.
+        * This might be racy...
+        */
+       if (!atomic_read(&num_events))
+               return NOTIFY_OK;
+
+       switch (action & ~CPU_TASKS_FROZEN) {
+       case CPU_ONLINE:
+       case CPU_ONLINE_FROZEN:
+               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 init_cpum_sampling_pmu(void)
+{
+       int err;
+
+       if (!cpum_sf_avail())
+               return -ENODEV;
+
+       sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
+       if (!sfdbg)
+               pr_err("Registering for s390dbf failed\n");
+       debug_register_view(sfdbg, &debug_sprintf_view);
+
+       err = register_external_interrupt(0x1407, cpumf_measurement_alert);
+       if (err) {
+               pr_err("Failed to register for CPU-measurement alerts\n");
+               goto out;
+       }
+
+       err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
+       if (err) {
+               pr_err("Failed to register cpum_sf pmu\n");
+               unregister_external_interrupt(0x1407, cpumf_measurement_alert);
+               goto out;
+       }
+       perf_cpu_notifier(cpumf_pmu_notifier);
+out:
+       return err;
+}
+arch_initcall(init_cpum_sampling_pmu);
index 4c1d336ce941b53031afff8d06d65775650f96f6..b9843ba9829f96c0013329ed84a92749d969a6ea 100644 (file)
@@ -16,6 +16,7 @@
 #include <linux/kvm_host.h>
 #include <linux/percpu.h>
 #include <linux/export.h>
+#include <linux/spinlock.h>
 #include <linux/sysfs.h>
 #include <asm/irq.h>
 #include <asm/cpu_mf.h>
@@ -36,6 +37,8 @@ int perf_num_counters(void)
 
        if (cpum_cf_avail())
                num += PERF_CPUM_CF_MAX_CTR;
+       if (cpum_sf_avail())
+               num += PERF_CPUM_SF_MAX_CTR;
 
        return num;
 }
@@ -93,24 +96,45 @@ unsigned long perf_misc_flags(struct pt_regs *regs)
                               : PERF_RECORD_MISC_KERNEL;
 }
 
-void perf_event_print_debug(void)
+void print_debug_cf(void)
 {
        struct cpumf_ctr_info cf_info;
-       unsigned long flags;
-       int cpu;
-
-       if (!cpum_cf_avail())
-               return;
-
-       local_irq_save(flags);
+       int cpu = smp_processor_id();
 
-       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);
+}
+
+static void print_debug_sf(void)
+{
+       struct hws_qsi_info_block si;
+       int cpu = smp_processor_id();
+
+       memset(&si, 0, sizeof(si));
+       if (qsi(&si)) {
+               pr_err("CPU[%i]: CPM_SF: qsi failed\n");
+               return;
+       }
+
+       pr_info("CPU[%i]: CPM_SF: as=%i es=%i cs=%i bsdes=%i dsdes=%i"
+               " min=%i max=%i cpu_speed=%i tear=%p dear=%p\n",
+               cpu, si.as, si.es, si.cs, si.bsdes, si.dsdes,
+               si.min_sampl_rate, si.max_sampl_rate, si.cpu_speed,
+               si.tear, si.dear);
+}
+
+void perf_event_print_debug(void)
+{
+       unsigned long flags;
 
+       local_irq_save(flags);
+       if (cpum_cf_avail())
+               print_debug_cf();
+       if (cpum_sf_avail())
+               print_debug_sf();
        local_irq_restore(flags);
 }