alpha: implement HW performance events on the EV67 and later CPUs
authorMichael Cree <mcree@orcon.net.nz>
Tue, 10 Aug 2010 00:20:08 +0000 (17:20 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 10 Aug 2010 03:45:04 +0000 (20:45 -0700)
This implements hardware performance events for the EV67 and later CPUs
within the Linux performance events subsystem.  Only using the performance
monitoring unit in HP/Compaq's so called "Aggregrate mode" is supported.

The code has been implemented in a manner that makes extension to other
older Alpha CPUs relatively straightforward should some mug wish to
indulge themselves.

Signed-off-by: Michael Cree <mcree@orcon.net.nz>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Jay Estabrook <jay.estabrook@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
arch/alpha/include/asm/perf_event.h
arch/alpha/kernel/Makefile
arch/alpha/kernel/irq_alpha.c
arch/alpha/kernel/perf_event.c [new file with mode: 0644]
arch/alpha/kernel/time.c

index 3bef8522017c23094e1cee52ffd2ef6da28a54e0..4157cd3c44a96d2811410b334a35f11923ba6fbf 100644 (file)
@@ -2,8 +2,14 @@
 #define __ASM_ALPHA_PERF_EVENT_H
 
 /* Alpha only supports software events through this interface. */
-static inline void set_perf_event_pending(void) { }
+extern void set_perf_event_pending(void);
 
 #define PERF_EVENT_INDEX_OFFSET 0
 
+#ifdef CONFIG_PERF_EVENTS
+extern void init_hw_perf_events(void);
+#else
+static inline void init_hw_perf_events(void)    { }
+#endif
+
 #endif /* __ASM_ALPHA_PERF_EVENT_H */
index 5a62fb46ef20b60759db3c7fdb68dad4daf3e525..1ee9b5b629b8fb37f707b678c2eeef5411869497 100644 (file)
@@ -15,6 +15,7 @@ obj-$(CONFIG_SMP)     += smp.o
 obj-$(CONFIG_PCI)      += pci.o pci_iommu.o pci-sysfs.o
 obj-$(CONFIG_SRM_ENV)  += srm_env.o
 obj-$(CONFIG_MODULES)  += module.o
+obj-$(CONFIG_PERF_EVENTS) += perf_event.o
 
 ifdef CONFIG_ALPHA_GENERIC
 
index cfde865b78e05f9c02c89d174887e38ef12590b2..5f77afb88e898b849f57e5c115f4688282e71c62 100644 (file)
@@ -10,6 +10,7 @@
 
 #include <asm/machvec.h>
 #include <asm/dma.h>
+#include <asm/perf_event.h>
 
 #include "proto.h"
 #include "irq_impl.h"
@@ -111,6 +112,8 @@ init_IRQ(void)
        wrent(entInt, 0);
 
        alpha_mv.init_irq();
+
+       init_hw_perf_events();
 }
 
 /*
diff --git a/arch/alpha/kernel/perf_event.c b/arch/alpha/kernel/perf_event.c
new file mode 100644 (file)
index 0000000..51c39fa
--- /dev/null
@@ -0,0 +1,842 @@
+/*
+ * Hardware performance events for the Alpha.
+ *
+ * We implement HW counts on the EV67 and subsequent CPUs only.
+ *
+ * (C) 2010 Michael J. Cree
+ *
+ * Somewhat based on the Sparc code, and to a lesser extent the PowerPC and
+ * ARM code, which are copyright by their respective authors.
+ */
+
+#include <linux/perf_event.h>
+#include <linux/kprobes.h>
+#include <linux/kernel.h>
+#include <linux/kdebug.h>
+#include <linux/mutex.h>
+
+#include <asm/hwrpb.h>
+#include <asm/atomic.h>
+#include <asm/irq.h>
+#include <asm/irq_regs.h>
+#include <asm/pal.h>
+#include <asm/wrperfmon.h>
+#include <asm/hw_irq.h>
+
+
+/* The maximum number of PMCs on any Alpha CPU whatsoever. */
+#define MAX_HWEVENTS 3
+#define PMC_NO_INDEX -1
+
+/* For tracking PMCs and the hw events they monitor on each CPU. */
+struct cpu_hw_events {
+       int                     enabled;
+       /* Number of events scheduled; also number entries valid in arrays below. */
+       int                     n_events;
+       /* Number events added since last hw_perf_disable(). */
+       int                     n_added;
+       /* Events currently scheduled. */
+       struct perf_event       *event[MAX_HWEVENTS];
+       /* Event type of each scheduled event. */
+       unsigned long           evtype[MAX_HWEVENTS];
+       /* Current index of each scheduled event; if not yet determined
+        * contains PMC_NO_INDEX.
+        */
+       int                     current_idx[MAX_HWEVENTS];
+       /* The active PMCs' config for easy use with wrperfmon(). */
+       unsigned long           config;
+       /* The active counters' indices for easy use with wrperfmon(). */
+       unsigned long           idx_mask;
+};
+DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
+
+
+
+/*
+ * A structure to hold the description of the PMCs available on a particular
+ * type of Alpha CPU.
+ */
+struct alpha_pmu_t {
+       /* Mapping of the perf system hw event types to indigenous event types */
+       const int *event_map;
+       /* The number of entries in the event_map */
+       int  max_events;
+       /* The number of PMCs on this Alpha */
+       int  num_pmcs;
+       /*
+        * All PMC counters reside in the IBOX register PCTR.  This is the
+        * LSB of the counter.
+        */
+       int  pmc_count_shift[MAX_HWEVENTS];
+       /*
+        * The mask that isolates the PMC bits when the LSB of the counter
+        * is shifted to bit 0.
+        */
+       unsigned long pmc_count_mask[MAX_HWEVENTS];
+       /* The maximum period the PMC can count. */
+       unsigned long pmc_max_period[MAX_HWEVENTS];
+       /*
+        * The maximum value that may be written to the counter due to
+        * hardware restrictions is pmc_max_period - pmc_left.
+        */
+       long pmc_left[3];
+        /* Subroutine for allocation of PMCs.  Enforces constraints. */
+       int (*check_constraints)(struct perf_event **, unsigned long *, int);
+};
+
+/*
+ * The Alpha CPU PMU description currently in operation.  This is set during
+ * the boot process to the specific CPU of the machine.
+ */
+static const struct alpha_pmu_t *alpha_pmu;
+
+
+#define HW_OP_UNSUPPORTED -1
+
+/*
+ * The hardware description of the EV67, EV68, EV69, EV7 and EV79 PMUs
+ * follow. Since they are identical we refer to them collectively as the
+ * EV67 henceforth.
+ */
+
+/*
+ * EV67 PMC event types
+ *
+ * There is no one-to-one mapping of the possible hw event types to the
+ * actual codes that are used to program the PMCs hence we introduce our
+ * own hw event type identifiers.
+ */
+enum ev67_pmc_event_type {
+       EV67_CYCLES = 1,
+       EV67_INSTRUCTIONS,
+       EV67_BCACHEMISS,
+       EV67_MBOXREPLAY,
+       EV67_LAST_ET
+};
+#define EV67_NUM_EVENT_TYPES (EV67_LAST_ET-EV67_CYCLES)
+
+
+/* Mapping of the hw event types to the perf tool interface */
+static const int ev67_perfmon_event_map[] = {
+       [PERF_COUNT_HW_CPU_CYCLES]       = EV67_CYCLES,
+       [PERF_COUNT_HW_INSTRUCTIONS]     = EV67_INSTRUCTIONS,
+       [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
+       [PERF_COUNT_HW_CACHE_MISSES]     = EV67_BCACHEMISS,
+};
+
+struct ev67_mapping_t {
+       int config;
+       int idx;
+};
+
+/*
+ * The mapping used for one event only - these must be in same order as enum
+ * ev67_pmc_event_type definition.
+ */
+static const struct ev67_mapping_t ev67_mapping[] = {
+       {EV67_PCTR_INSTR_CYCLES, 1},     /* EV67_CYCLES, */
+       {EV67_PCTR_INSTR_CYCLES, 0},     /* EV67_INSTRUCTIONS */
+       {EV67_PCTR_INSTR_BCACHEMISS, 1}, /* EV67_BCACHEMISS */
+       {EV67_PCTR_CYCLES_MBOX, 1}       /* EV67_MBOXREPLAY */
+};
+
+
+/*
+ * Check that a group of events can be simultaneously scheduled on to the
+ * EV67 PMU.  Also allocate counter indices and config.
+ */
+static int ev67_check_constraints(struct perf_event **event,
+                               unsigned long *evtype, int n_ev)
+{
+       int idx0;
+       unsigned long config;
+
+       idx0 = ev67_mapping[evtype[0]-1].idx;
+       config = ev67_mapping[evtype[0]-1].config;
+       if (n_ev == 1)
+               goto success;
+
+       BUG_ON(n_ev != 2);
+
+       if (evtype[0] == EV67_MBOXREPLAY || evtype[1] == EV67_MBOXREPLAY) {
+               /* MBOX replay traps must be on PMC 1 */
+               idx0 = (evtype[0] == EV67_MBOXREPLAY) ? 1 : 0;
+               /* Only cycles can accompany MBOX replay traps */
+               if (evtype[idx0] == EV67_CYCLES) {
+                       config = EV67_PCTR_CYCLES_MBOX;
+                       goto success;
+               }
+       }
+
+       if (evtype[0] == EV67_BCACHEMISS || evtype[1] == EV67_BCACHEMISS) {
+               /* Bcache misses must be on PMC 1 */
+               idx0 = (evtype[0] == EV67_BCACHEMISS) ? 1 : 0;
+               /* Only instructions can accompany Bcache misses */
+               if (evtype[idx0] == EV67_INSTRUCTIONS) {
+                       config = EV67_PCTR_INSTR_BCACHEMISS;
+                       goto success;
+               }
+       }
+
+       if (evtype[0] == EV67_INSTRUCTIONS || evtype[1] == EV67_INSTRUCTIONS) {
+               /* Instructions must be on PMC 0 */
+               idx0 = (evtype[0] == EV67_INSTRUCTIONS) ? 0 : 1;
+               /* By this point only cycles can accompany instructions */
+               if (evtype[idx0^1] == EV67_CYCLES) {
+                       config = EV67_PCTR_INSTR_CYCLES;
+                       goto success;
+               }
+       }
+
+       /* Otherwise, darn it, there is a conflict.  */
+       return -1;
+
+success:
+       event[0]->hw.idx = idx0;
+       event[0]->hw.config_base = config;
+       if (n_ev == 2) {
+               event[1]->hw.idx = idx0 ^ 1;
+               event[1]->hw.config_base = config;
+       }
+       return 0;
+}
+
+
+static const struct alpha_pmu_t ev67_pmu = {
+       .event_map = ev67_perfmon_event_map,
+       .max_events = ARRAY_SIZE(ev67_perfmon_event_map),
+       .num_pmcs = 2,
+       .pmc_count_shift = {EV67_PCTR_0_COUNT_SHIFT, EV67_PCTR_1_COUNT_SHIFT, 0},
+       .pmc_count_mask = {EV67_PCTR_0_COUNT_MASK,  EV67_PCTR_1_COUNT_MASK,  0},
+       .pmc_max_period = {(1UL<<20) - 1, (1UL<<20) - 1, 0},
+       .pmc_left = {16, 4, 0},
+       .check_constraints = ev67_check_constraints
+};
+
+
+
+/*
+ * Helper routines to ensure that we read/write only the correct PMC bits
+ * when calling the wrperfmon PALcall.
+ */
+static inline void alpha_write_pmc(int idx, unsigned long val)
+{
+       val &= alpha_pmu->pmc_count_mask[idx];
+       val <<= alpha_pmu->pmc_count_shift[idx];
+       val |= (1<<idx);
+       wrperfmon(PERFMON_CMD_WRITE, val);
+}
+
+static inline unsigned long alpha_read_pmc(int idx)
+{
+       unsigned long val;
+
+       val = wrperfmon(PERFMON_CMD_READ, 0);
+       val >>= alpha_pmu->pmc_count_shift[idx];
+       val &= alpha_pmu->pmc_count_mask[idx];
+       return val;
+}
+
+/* Set a new period to sample over */
+static int alpha_perf_event_set_period(struct perf_event *event,
+                               struct hw_perf_event *hwc, int idx)
+{
+       long left = atomic64_read(&hwc->period_left);
+       long period = hwc->sample_period;
+       int ret = 0;
+
+       if (unlikely(left <= -period)) {
+               left = period;
+               atomic64_set(&hwc->period_left, left);
+               hwc->last_period = period;
+               ret = 1;
+       }
+
+       if (unlikely(left <= 0)) {
+               left += period;
+               atomic64_set(&hwc->period_left, left);
+               hwc->last_period = period;
+               ret = 1;
+       }
+
+       /*
+        * Hardware restrictions require that the counters must not be
+        * written with values that are too close to the maximum period.
+        */
+       if (unlikely(left < alpha_pmu->pmc_left[idx]))
+               left = alpha_pmu->pmc_left[idx];
+
+       if (left > (long)alpha_pmu->pmc_max_period[idx])
+               left = alpha_pmu->pmc_max_period[idx];
+
+       atomic64_set(&hwc->prev_count, (unsigned long)(-left));
+
+       alpha_write_pmc(idx, (unsigned long)(-left));
+
+       perf_event_update_userpage(event);
+
+       return ret;
+}
+
+
+/*
+ * Calculates the count (the 'delta') since the last time the PMC was read.
+ *
+ * As the PMCs' full period can easily be exceeded within the perf system
+ * sampling period we cannot use any high order bits as a guard bit in the
+ * PMCs to detect overflow as is done by other architectures.  The code here
+ * calculates the delta on the basis that there is no overflow when ovf is
+ * zero.  The value passed via ovf by the interrupt handler corrects for
+ * overflow.
+ *
+ * This can be racey on rare occasions -- a call to this routine can occur
+ * with an overflowed counter just before the PMI service routine is called.
+ * The check for delta negative hopefully always rectifies this situation.
+ */
+static unsigned long alpha_perf_event_update(struct perf_event *event,
+                                       struct hw_perf_event *hwc, int idx, long ovf)
+{
+       long prev_raw_count, new_raw_count;
+       long delta;
+
+again:
+       prev_raw_count = atomic64_read(&hwc->prev_count);
+       new_raw_count = alpha_read_pmc(idx);
+
+       if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
+                            new_raw_count) != prev_raw_count)
+               goto again;
+
+       delta = (new_raw_count  - (prev_raw_count & alpha_pmu->pmc_count_mask[idx])) + ovf;
+
+       /* It is possible on very rare occasions that the PMC has overflowed
+        * but the interrupt is yet to come.  Detect and fix this situation.
+        */
+       if (unlikely(delta < 0)) {
+               delta += alpha_pmu->pmc_max_period[idx] + 1;
+       }
+
+       atomic64_add(delta, &event->count);
+       atomic64_sub(delta, &hwc->period_left);
+
+       return new_raw_count;
+}
+
+
+/*
+ * Collect all HW events into the array event[].
+ */
+static int collect_events(struct perf_event *group, int max_count,
+                         struct perf_event *event[], unsigned long *evtype,
+                         int *current_idx)
+{
+       struct perf_event *pe;
+       int n = 0;
+
+       if (!is_software_event(group)) {
+               if (n >= max_count)
+                       return -1;
+               event[n] = group;
+               evtype[n] = group->hw.event_base;
+               current_idx[n++] = PMC_NO_INDEX;
+       }
+       list_for_each_entry(pe, &group->sibling_list, group_entry) {
+               if (!is_software_event(pe) && pe->state != PERF_EVENT_STATE_OFF) {
+                       if (n >= max_count)
+                               return -1;
+                       event[n] = pe;
+                       evtype[n] = pe->hw.event_base;
+                       current_idx[n++] = PMC_NO_INDEX;
+               }
+       }
+       return n;
+}
+
+
+
+/*
+ * Check that a group of events can be simultaneously scheduled on to the PMU.
+ */
+static int alpha_check_constraints(struct perf_event **events,
+                                  unsigned long *evtypes, int n_ev)
+{
+
+       /* No HW events is possible from hw_perf_group_sched_in(). */
+       if (n_ev == 0)
+               return 0;
+
+       if (n_ev > alpha_pmu->num_pmcs)
+               return -1;
+
+       return alpha_pmu->check_constraints(events, evtypes, n_ev);
+}
+
+
+/*
+ * If new events have been scheduled then update cpuc with the new
+ * configuration.  This may involve shifting cycle counts from one PMC to
+ * another.
+ */
+static void maybe_change_configuration(struct cpu_hw_events *cpuc)
+{
+       int j;
+
+       if (cpuc->n_added == 0)
+               return;
+
+       /* Find counters that are moving to another PMC and update */
+       for (j = 0; j < cpuc->n_events; j++) {
+               struct perf_event *pe = cpuc->event[j];
+
+               if (cpuc->current_idx[j] != PMC_NO_INDEX &&
+                       cpuc->current_idx[j] != pe->hw.idx) {
+                       alpha_perf_event_update(pe, &pe->hw, cpuc->current_idx[j], 0);
+                       cpuc->current_idx[j] = PMC_NO_INDEX;
+               }
+       }
+
+       /* Assign to counters all unassigned events. */
+       cpuc->idx_mask = 0;
+       for (j = 0; j < cpuc->n_events; j++) {
+               struct perf_event *pe = cpuc->event[j];
+               struct hw_perf_event *hwc = &pe->hw;
+               int idx = hwc->idx;
+
+               if (cpuc->current_idx[j] != PMC_NO_INDEX) {
+                       cpuc->idx_mask |= (1<<cpuc->current_idx[j]);
+                       continue;
+               }
+
+               alpha_perf_event_set_period(pe, hwc, idx);
+               cpuc->current_idx[j] = idx;
+               cpuc->idx_mask |= (1<<cpuc->current_idx[j]);
+       }
+       cpuc->config = cpuc->event[0]->hw.config_base;
+}
+
+
+
+/* Schedule perf HW event on to PMU.
+ *  - this function is called from outside this module via the pmu struct
+ *    returned from perf event initialisation.
+ */
+static int alpha_pmu_enable(struct perf_event *event)
+{
+       struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+       int n0;
+       int ret;
+       unsigned long flags;
+
+       /*
+        * The Sparc code has the IRQ disable first followed by the perf
+        * disable, however this can lead to an overflowed counter with the
+        * PMI disabled on rare occasions.  The alpha_perf_event_update()
+        * routine should detect this situation by noting a negative delta,
+        * nevertheless we disable the PMCs first to enable a potential
+        * final PMI to occur before we disable interrupts.
+        */
+       perf_disable();
+       local_irq_save(flags);
+
+       /* Default to error to be returned */
+       ret = -EAGAIN;
+
+       /* Insert event on to PMU and if successful modify ret to valid return */
+       n0 = cpuc->n_events;
+       if (n0 < alpha_pmu->num_pmcs) {
+               cpuc->event[n0] = event;
+               cpuc->evtype[n0] = event->hw.event_base;
+               cpuc->current_idx[n0] = PMC_NO_INDEX;
+
+               if (!alpha_check_constraints(cpuc->event, cpuc->evtype, n0+1)) {
+                       cpuc->n_events++;
+                       cpuc->n_added++;
+                       ret = 0;
+               }
+       }
+
+       local_irq_restore(flags);
+       perf_enable();
+
+       return ret;
+}
+
+
+
+/* Disable performance monitoring unit
+ *  - this function is called from outside this module via the pmu struct
+ *    returned from perf event initialisation.
+ */
+static void alpha_pmu_disable(struct perf_event *event)
+{
+       struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+       struct hw_perf_event *hwc = &event->hw;
+       unsigned long flags;
+       int j;
+
+       perf_disable();
+       local_irq_save(flags);
+
+       for (j = 0; j < cpuc->n_events; j++) {
+               if (event == cpuc->event[j]) {
+                       int idx = cpuc->current_idx[j];
+
+                       /* Shift remaining entries down into the existing
+                        * slot.
+                        */
+                       while (++j < cpuc->n_events) {
+                               cpuc->event[j - 1] = cpuc->event[j];
+                               cpuc->evtype[j - 1] = cpuc->evtype[j];
+                               cpuc->current_idx[j - 1] =
+                                       cpuc->current_idx[j];
+                       }
+
+                       /* Absorb the final count and turn off the event. */
+                       alpha_perf_event_update(event, hwc, idx, 0);
+                       perf_event_update_userpage(event);
+
+                       cpuc->idx_mask &= ~(1UL<<idx);
+                       cpuc->n_events--;
+                       break;
+               }
+       }
+
+       local_irq_restore(flags);
+       perf_enable();
+}
+
+
+static void alpha_pmu_read(struct perf_event *event)
+{
+       struct hw_perf_event *hwc = &event->hw;
+
+       alpha_perf_event_update(event, hwc, hwc->idx, 0);
+}
+
+
+static void alpha_pmu_unthrottle(struct perf_event *event)
+{
+       struct hw_perf_event *hwc = &event->hw;
+       struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+
+       cpuc->idx_mask |= 1UL<<hwc->idx;
+       wrperfmon(PERFMON_CMD_ENABLE, (1UL<<hwc->idx));
+}
+
+
+/*
+ * Check that CPU performance counters are supported.
+ * - currently support EV67 and later CPUs.
+ * - actually some later revisions of the EV6 have the same PMC model as the
+ *     EV67 but we don't do suffiently deep CPU detection to detect them.
+ *     Bad luck to the very few people who might have one, I guess.
+ */
+static int supported_cpu(void)
+{
+       struct percpu_struct *cpu;
+       unsigned long cputype;
+
+       /* Get cpu type from HW */
+       cpu = (struct percpu_struct *)((char *)hwrpb + hwrpb->processor_offset);
+       cputype = cpu->type & 0xffffffff;
+       /* Include all of EV67, EV68, EV7, EV79 and EV69 as supported. */
+       return (cputype >= EV67_CPU) && (cputype <= EV69_CPU);
+}
+
+
+
+static void hw_perf_event_destroy(struct perf_event *event)
+{
+       /* Nothing to be done! */
+       return;
+}
+
+
+
+static int __hw_perf_event_init(struct perf_event *event)
+{
+       struct perf_event_attr *attr = &event->attr;
+       struct hw_perf_event *hwc = &event->hw;
+       struct perf_event *evts[MAX_HWEVENTS];
+       unsigned long evtypes[MAX_HWEVENTS];
+       int idx_rubbish_bin[MAX_HWEVENTS];
+       int ev;
+       int n;
+
+       /* We only support a limited range of HARDWARE event types with one
+        * only programmable via a RAW event type.
+        */
+       if (attr->type == PERF_TYPE_HARDWARE) {
+               if (attr->config >= alpha_pmu->max_events)
+                       return -EINVAL;
+               ev = alpha_pmu->event_map[attr->config];
+       } else if (attr->type == PERF_TYPE_HW_CACHE) {
+               return -EOPNOTSUPP;
+       } else if (attr->type == PERF_TYPE_RAW) {
+               ev = attr->config & 0xff;
+       } else {
+               return -EOPNOTSUPP;
+       }
+
+       if (ev < 0) {
+               return ev;
+       }
+
+       /* The EV67 does not support mode exclusion */
+       if (attr->exclude_kernel || attr->exclude_user
+                       || attr->exclude_hv || attr->exclude_idle) {
+               return -EPERM;
+       }
+
+       /*
+        * We place the event type in event_base here and leave calculation
+        * of the codes to programme the PMU for alpha_pmu_enable() because
+        * it is only then we will know what HW events are actually
+        * scheduled on to the PMU.  At that point the code to programme the
+        * PMU is put into config_base and the PMC to use is placed into
+        * idx.  We initialise idx (below) to PMC_NO_INDEX to indicate that
+        * it is yet to be determined.
+        */
+       hwc->event_base = ev;
+
+       /* Collect events in a group together suitable for calling
+        * alpha_check_constraints() to verify that the group as a whole can
+        * be scheduled on to the PMU.
+        */
+       n = 0;
+       if (event->group_leader != event) {
+               n = collect_events(event->group_leader,
+                               alpha_pmu->num_pmcs - 1,
+                               evts, evtypes, idx_rubbish_bin);
+               if (n < 0)
+                       return -EINVAL;
+       }
+       evtypes[n] = hwc->event_base;
+       evts[n] = event;
+
+       if (alpha_check_constraints(evts, evtypes, n + 1))
+               return -EINVAL;
+
+       /* Indicate that PMU config and idx are yet to be determined. */
+       hwc->config_base = 0;
+       hwc->idx = PMC_NO_INDEX;
+
+       event->destroy = hw_perf_event_destroy;
+
+       /*
+        * Most architectures reserve the PMU for their use at this point.
+        * As there is no existing mechanism to arbitrate usage and there
+        * appears to be no other user of the Alpha PMU we just assume
+        * that we can just use it, hence a NO-OP here.
+        *
+        * Maybe an alpha_reserve_pmu() routine should be implemented but is
+        * anything else ever going to use it?
+        */
+
+       if (!hwc->sample_period) {
+               hwc->sample_period = alpha_pmu->pmc_max_period[0];
+               hwc->last_period = hwc->sample_period;
+               atomic64_set(&hwc->period_left, hwc->sample_period);
+       }
+
+       return 0;
+}
+
+static const struct pmu pmu = {
+       .enable         = alpha_pmu_enable,
+       .disable        = alpha_pmu_disable,
+       .read           = alpha_pmu_read,
+       .unthrottle     = alpha_pmu_unthrottle,
+};
+
+
+/*
+ * Main entry point to initialise a HW performance event.
+ */
+const struct pmu *hw_perf_event_init(struct perf_event *event)
+{
+       int err;
+
+       if (!alpha_pmu)
+               return ERR_PTR(-ENODEV);
+
+       /* Do the real initialisation work. */
+       err = __hw_perf_event_init(event);
+
+       if (err)
+               return ERR_PTR(err);
+
+       return &pmu;
+}
+
+
+
+/*
+ * Main entry point - enable HW performance counters.
+ */
+void hw_perf_enable(void)
+{
+       struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+
+       if (cpuc->enabled)
+               return;
+
+       cpuc->enabled = 1;
+       barrier();
+
+       if (cpuc->n_events > 0) {
+               /* Update cpuc with information from any new scheduled events. */
+               maybe_change_configuration(cpuc);
+
+               /* Start counting the desired events. */
+               wrperfmon(PERFMON_CMD_LOGGING_OPTIONS, EV67_PCTR_MODE_AGGREGATE);
+               wrperfmon(PERFMON_CMD_DESIRED_EVENTS, cpuc->config);
+               wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask);
+       }
+}
+
+
+/*
+ * Main entry point - disable HW performance counters.
+ */
+
+void hw_perf_disable(void)
+{
+       struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+
+       if (!cpuc->enabled)
+               return;
+
+       cpuc->enabled = 0;
+       cpuc->n_added = 0;
+
+       wrperfmon(PERFMON_CMD_DISABLE, cpuc->idx_mask);
+}
+
+
+/*
+ * Main entry point - don't know when this is called but it
+ * obviously dumps debug info.
+ */
+void perf_event_print_debug(void)
+{
+       unsigned long flags;
+       unsigned long pcr;
+       int pcr0, pcr1;
+       int cpu;
+
+       if (!supported_cpu())
+               return;
+
+       local_irq_save(flags);
+
+       cpu = smp_processor_id();
+
+       pcr = wrperfmon(PERFMON_CMD_READ, 0);
+       pcr0 = (pcr >> alpha_pmu->pmc_count_shift[0]) & alpha_pmu->pmc_count_mask[0];
+       pcr1 = (pcr >> alpha_pmu->pmc_count_shift[1]) & alpha_pmu->pmc_count_mask[1];
+
+       pr_info("CPU#%d: PCTR0[%06x] PCTR1[%06x]\n", cpu, pcr0, pcr1);
+
+       local_irq_restore(flags);
+}
+
+
+/*
+ * Performance Monitoring Interrupt Service Routine called when a PMC
+ * overflows.  The PMC that overflowed is passed in la_ptr.
+ */
+static void alpha_perf_event_irq_handler(unsigned long la_ptr,
+                                       struct pt_regs *regs)
+{
+       struct cpu_hw_events *cpuc;
+       struct perf_sample_data data;
+       struct perf_event *event;
+       struct hw_perf_event *hwc;
+       int idx, j;
+
+       __get_cpu_var(irq_pmi_count)++;
+       cpuc = &__get_cpu_var(cpu_hw_events);
+
+       /* Completely counting through the PMC's period to trigger a new PMC
+        * overflow interrupt while in this interrupt routine is utterly
+        * disastrous!  The EV6 and EV67 counters are sufficiently large to
+        * prevent this but to be really sure disable the PMCs.
+        */
+       wrperfmon(PERFMON_CMD_DISABLE, cpuc->idx_mask);
+
+       /* la_ptr is the counter that overflowed. */
+       if (unlikely(la_ptr >= perf_max_events)) {
+               /* This should never occur! */
+               irq_err_count++;
+               pr_warning("PMI: silly index %ld\n", la_ptr);
+               wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask);
+               return;
+       }
+
+       idx = la_ptr;
+
+       perf_sample_data_init(&data, 0);
+       for (j = 0; j < cpuc->n_events; j++) {
+               if (cpuc->current_idx[j] == idx)
+                       break;
+       }
+
+       if (unlikely(j == cpuc->n_events)) {
+               /* This can occur if the event is disabled right on a PMC overflow. */
+               wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask);
+               return;
+       }
+
+       event = cpuc->event[j];
+
+       if (unlikely(!event)) {
+               /* This should never occur! */
+               irq_err_count++;
+               pr_warning("PMI: No event at index %d!\n", idx);
+               wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask);
+               return;
+       }
+
+       hwc = &event->hw;
+       alpha_perf_event_update(event, hwc, idx, alpha_pmu->pmc_max_period[idx]+1);
+       data.period = event->hw.last_period;
+
+       if (alpha_perf_event_set_period(event, hwc, idx)) {
+               if (perf_event_overflow(event, 1, &data, regs)) {
+                       /* Interrupts coming too quickly; "throttle" the
+                        * counter, i.e., disable it for a little while.
+                        */
+                       cpuc->idx_mask &= ~(1UL<<idx);
+               }
+       }
+       wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask);
+
+       return;
+}
+
+
+
+/*
+ * Init call to initialise performance events at kernel startup.
+ */
+void __init init_hw_perf_events(void)
+{
+       pr_info("Performance events: ");
+
+       if (!supported_cpu()) {
+               pr_cont("No support for your CPU.\n");
+               return;
+       }
+
+       pr_cont("Supported CPU type!\n");
+
+       /* Override performance counter IRQ vector */
+
+       perf_irq = alpha_perf_event_irq_handler;
+
+       /* And set up PMU specification */
+       alpha_pmu = &ev67_pmu;
+       perf_max_events = alpha_pmu->num_pmcs;
+}
+
index 1efbed82c0fd38a2bb0ecd925d80bef884801642..eacceb26d9c8aa8bd250b409e821c277a9b51c73 100644 (file)
@@ -41,6 +41,7 @@
 #include <linux/init.h>
 #include <linux/bcd.h>
 #include <linux/profile.h>
+#include <linux/perf_event.h>
 
 #include <asm/uaccess.h>
 #include <asm/io.h>
@@ -82,6 +83,26 @@ static struct {
 
 unsigned long est_cycle_freq;
 
+#ifdef CONFIG_PERF_EVENTS
+
+DEFINE_PER_CPU(u8, perf_event_pending);
+
+#define set_perf_event_pending_flag()  __get_cpu_var(perf_event_pending) = 1
+#define test_perf_event_pending()      __get_cpu_var(perf_event_pending)
+#define clear_perf_event_pending()     __get_cpu_var(perf_event_pending) = 0
+
+void set_perf_event_pending(void)
+{
+       set_perf_event_pending_flag();
+}
+
+#else  /* CONFIG_PERF_EVENTS */
+
+#define test_perf_event_pending()      0
+#define clear_perf_event_pending()
+
+#endif /* CONFIG_PERF_EVENTS */
+
 
 static inline __u32 rpcc(void)
 {
@@ -175,6 +196,11 @@ irqreturn_t timer_interrupt(int irq, void *dev)
                update_process_times(user_mode(get_irq_regs()));
 #endif
 
+       if (test_perf_event_pending()) {
+               clear_perf_event_pending();
+               perf_event_do_pending();
+       }
+
        return IRQ_HANDLED;
 }