]> pilppa.com Git - linux-2.6-omap-h63xx.git/commitdiff
[POWERPC] cell: Add oprofile support
authorMaynard Johnson <maynardj@us.ibm.com>
Mon, 20 Nov 2006 17:45:16 +0000 (18:45 +0100)
committerPaul Mackerras <paulus@samba.org>
Mon, 4 Dec 2006 09:40:14 +0000 (20:40 +1100)
Add PPU event-based and cycle-based profiling support to Oprofile for Cell.

Oprofile is expected to collect data on all CPUs simultaneously.
However, there is one set of performance counters per node.  There are
two hardware threads or virtual CPUs on each node.  Hence, OProfile must
multiplex in time the performance counter collection on the two virtual
CPUs.

The multiplexing of the performance counters is done by a virtual
counter routine.  Initially, the counters are configured to collect data
on the even CPUs in the system, one CPU per node.  In order to capture
the PC for the virtual CPU when the performance counter interrupt occurs
(the specified number of events between samples has occurred), the even
processors are configured to handle the performance counter interrupts
for their node.  The virtual counter routine is called via a kernel
timer after the virtual sample time.  The routine stops the counters,
saves the current counts, loads the last counts for the other virtual
CPU on the node, sets interrupts to be handled by the other virtual CPU
and restarts the counters, the virtual timer routine is scheduled to run
again.  The virtual sample time is kept relatively small to make sure
sampling occurs on both CPUs on the node with a relatively small
granularity.  Whenever the counters overflow, the performance counter
interrupt is called to collect the PC for the CPU where data is being
collected.

The oprofile driver relies on a firmware RTAS call to setup the debug bus
to route the desired signals to the performance counter hardware to be
counted.  The RTAS call must set the routing registers appropriately in
each of the islands to pass the signals down the debug bus as well as
routing the signals from a particular island onto the bus.  There is a
second firmware RTAS call to reset the debug bus to the non pass thru
state when the counters are not in use.

Signed-off-by: Carl Love <carll@us.ibm.com>
Signed-off-by: Maynard Johnson <mpjohn@us.ibm.com>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
arch/powerpc/configs/cell_defconfig
arch/powerpc/kernel/cputable.c
arch/powerpc/oprofile/Makefile
arch/powerpc/oprofile/common.c
arch/powerpc/oprofile/op_model_cell.c [new file with mode: 0644]
arch/powerpc/platforms/cell/cbe_regs.c
arch/powerpc/platforms/cell/pmu.c
include/asm-powerpc/cell-pmu.h
include/asm-powerpc/cputable.h
include/asm-powerpc/oprofile_impl.h

index dbcd4e3329d04f320ebf6b00306438c112425e3e..3d6a2f10a7dd612fb167771b69c4535fe3c1ff87 100644 (file)
@@ -1121,7 +1121,8 @@ CONFIG_PLIST=y
 #
 # Instrumentation Support
 #
-# CONFIG_PROFILING is not set
+CONFIG_PROFILING=y
+CONFIG_OPROFILE=y
 # CONFIG_KPROBES is not set
 
 #
index 1e4ed0731d15d02dab603f7228474219db20efdd..992121b2d261614d88d25afd2282298ca49bca49 100644 (file)
@@ -304,6 +304,9 @@ static struct cpu_spec cpu_specs[] = {
                        PPC_FEATURE_SMT,
                .icache_bsize           = 128,
                .dcache_bsize           = 128,
+               .num_pmcs               = 4,
+               .oprofile_cpu_type      = "ppc64/cell-be",
+               .oprofile_type          = PPC_OPROFILE_CELL,
                .platform               = "ppc-cell-be",
        },
        {       /* PA Semi PA6T */
index 0b5df9c96ae0a4d1560724e4ddf9d7ca2d1b1f37..51c510fed7f7b4a259362dee4feb1a35c98bbf69 100644 (file)
@@ -11,6 +11,7 @@ DRIVER_OBJS := $(addprefix ../../../drivers/oprofile/, \
                timer_int.o )
 
 oprofile-y := $(DRIVER_OBJS) common.o backtrace.o
+oprofile-$(CONFIG_PPC_CELL) += op_model_cell.o
 oprofile-$(CONFIG_PPC64) += op_model_rs64.o op_model_power4.o
 oprofile-$(CONFIG_FSL_BOOKE) += op_model_fsl_booke.o
 oprofile-$(CONFIG_6xx) += op_model_7450.o
index 63bbef3b63f18be04701394cb17b255d1c42b20c..7a423437977ce5e43d20ce08bf142f4f8868ca77 100644 (file)
@@ -69,7 +69,10 @@ static void op_powerpc_cpu_start(void *dummy)
 
 static int op_powerpc_start(void)
 {
-       on_each_cpu(op_powerpc_cpu_start, NULL, 0, 1);
+       if (model->global_start)
+               model->global_start(ctr);
+       if (model->start)
+               on_each_cpu(op_powerpc_cpu_start, NULL, 0, 1);
        return 0;
 }
 
@@ -80,7 +83,10 @@ static inline void op_powerpc_cpu_stop(void *dummy)
 
 static void op_powerpc_stop(void)
 {
-       on_each_cpu(op_powerpc_cpu_stop, NULL, 0, 1);
+       if (model->stop)
+               on_each_cpu(op_powerpc_cpu_stop, NULL, 0, 1);
+        if (model->global_stop)
+                model->global_stop();
 }
 
 static int op_powerpc_create_files(struct super_block *sb, struct dentry *root)
@@ -141,6 +147,11 @@ int __init oprofile_arch_init(struct oprofile_operations *ops)
 
        switch (cur_cpu_spec->oprofile_type) {
 #ifdef CONFIG_PPC64
+#ifdef CONFIG_PPC_CELL
+               case PPC_OPROFILE_CELL:
+                       model = &op_model_cell;
+                       break;
+#endif
                case PPC_OPROFILE_RS64:
                        model = &op_model_rs64;
                        break;
diff --git a/arch/powerpc/oprofile/op_model_cell.c b/arch/powerpc/oprofile/op_model_cell.c
new file mode 100644 (file)
index 0000000..2eb15f3
--- /dev/null
@@ -0,0 +1,724 @@
+/*
+ * Cell Broadband Engine OProfile Support
+ *
+ * (C) Copyright IBM Corporation 2006
+ *
+ * Author: David Erb (djerb@us.ibm.com)
+ * Modifications:
+ *         Carl Love <carll@us.ibm.com>
+ *         Maynard Johnson <maynardj@us.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+#include <linux/cpufreq.h>
+#include <linux/delay.h>
+#include <linux/init.h>
+#include <linux/jiffies.h>
+#include <linux/kthread.h>
+#include <linux/oprofile.h>
+#include <linux/percpu.h>
+#include <linux/smp.h>
+#include <linux/spinlock.h>
+#include <linux/timer.h>
+#include <asm/cell-pmu.h>
+#include <asm/cputable.h>
+#include <asm/firmware.h>
+#include <asm/io.h>
+#include <asm/oprofile_impl.h>
+#include <asm/processor.h>
+#include <asm/prom.h>
+#include <asm/ptrace.h>
+#include <asm/reg.h>
+#include <asm/rtas.h>
+#include <asm/system.h>
+
+#include "../platforms/cell/interrupt.h"
+
+#define PPU_CYCLES_EVENT_NUM 1 /*  event number for CYCLES */
+#define CBE_COUNT_ALL_CYCLES 0x42800000        /* PPU cycle event specifier */
+
+#define NUM_THREADS 2
+#define VIRT_CNTR_SW_TIME_NS 100000000 // 0.5 seconds
+
+struct pmc_cntrl_data {
+       unsigned long vcntr;
+       unsigned long evnts;
+       unsigned long masks;
+       unsigned long enabled;
+};
+
+/*
+ * ibm,cbe-perftools rtas parameters
+ */
+
+struct pm_signal {
+       u16 cpu;                /* Processor to modify */
+       u16 sub_unit;           /* hw subunit this applies to (if applicable) */
+       u16 signal_group;       /* Signal Group to Enable/Disable */
+       u8 bus_word;            /* Enable/Disable on this Trace/Trigger/Event
+                                * Bus Word(s) (bitmask)
+                                */
+       u8 bit;                 /* Trigger/Event bit (if applicable) */
+};
+
+/*
+ * rtas call arguments
+ */
+enum {
+       SUBFUNC_RESET = 1,
+       SUBFUNC_ACTIVATE = 2,
+       SUBFUNC_DEACTIVATE = 3,
+
+       PASSTHRU_IGNORE = 0,
+       PASSTHRU_ENABLE = 1,
+       PASSTHRU_DISABLE = 2,
+};
+
+struct pm_cntrl {
+       u16 enable;
+       u16 stop_at_max;
+       u16 trace_mode;
+       u16 freeze;
+       u16 count_mode;
+};
+
+static struct {
+       u32 group_control;
+       u32 debug_bus_control;
+       struct pm_cntrl pm_cntrl;
+       u32 pm07_cntrl[NR_PHYS_CTRS];
+} pm_regs;
+
+
+#define GET_SUB_UNIT(x) ((x & 0x0000f000) >> 12)
+#define GET_BUS_WORD(x) ((x & 0x000000f0) >> 4)
+#define GET_BUS_TYPE(x) ((x & 0x00000300) >> 8)
+#define GET_POLARITY(x) ((x & 0x00000002) >> 1)
+#define GET_COUNT_CYCLES(x) (x & 0x00000001)
+#define GET_INPUT_CONTROL(x) ((x & 0x00000004) >> 2)
+
+
+static DEFINE_PER_CPU(unsigned long[NR_PHYS_CTRS], pmc_values);
+
+static struct pmc_cntrl_data pmc_cntrl[NUM_THREADS][NR_PHYS_CTRS];
+
+/* Interpetation of hdw_thread:
+ * 0 - even virtual cpus 0, 2, 4,...
+ * 1 - odd virtual cpus 1, 3, 5, ...
+ */
+static u32 hdw_thread;
+
+static u32 virt_cntr_inter_mask;
+static struct timer_list timer_virt_cntr;
+
+/* pm_signal needs to be global since it is initialized in
+ * cell_reg_setup at the time when the necessary information
+ * is available.
+ */
+static struct pm_signal pm_signal[NR_PHYS_CTRS];
+static int pm_rtas_token;
+
+static u32 reset_value[NR_PHYS_CTRS];
+static int num_counters;
+static int oprofile_running;
+static spinlock_t virt_cntr_lock = SPIN_LOCK_UNLOCKED;
+
+static u32 ctr_enabled;
+
+static unsigned char trace_bus[4];
+static unsigned char input_bus[2];
+
+/*
+ * Firmware interface functions
+ */
+static int
+rtas_ibm_cbe_perftools(int subfunc, int passthru,
+                      void *address, unsigned long length)
+{
+       u64 paddr = __pa(address);
+
+       return rtas_call(pm_rtas_token, 5, 1, NULL, subfunc, passthru,
+                        paddr >> 32, paddr & 0xffffffff, length);
+}
+
+static void pm_rtas_reset_signals(u32 node)
+{
+       int ret;
+       struct pm_signal pm_signal_local;
+
+       /*  The debug bus is being set to the passthru disable state.
+        *  However, the FW still expects atleast one legal signal routing
+        *  entry or it will return an error on the arguments.  If we don't
+        *  supply a valid entry, we must ignore all return values.  Ignoring
+        *  all return values means we might miss an error we should be
+        *  concerned about.
+        */
+
+       /*  fw expects physical cpu #. */
+       pm_signal_local.cpu = node;
+       pm_signal_local.signal_group = 21;
+       pm_signal_local.bus_word = 1;
+       pm_signal_local.sub_unit = 0;
+       pm_signal_local.bit = 0;
+
+       ret = rtas_ibm_cbe_perftools(SUBFUNC_RESET, PASSTHRU_DISABLE,
+                                    &pm_signal_local,
+                                    sizeof(struct pm_signal));
+
+       if (ret)
+               printk(KERN_WARNING "%s: rtas returned: %d\n",
+                      __FUNCTION__, ret);
+}
+
+static void pm_rtas_activate_signals(u32 node, u32 count)
+{
+       int ret;
+       int j;
+       struct pm_signal pm_signal_local[NR_PHYS_CTRS];
+
+       for (j = 0; j < count; j++) {
+               /* fw expects physical cpu # */
+               pm_signal_local[j].cpu = node;
+               pm_signal_local[j].signal_group = pm_signal[j].signal_group;
+               pm_signal_local[j].bus_word = pm_signal[j].bus_word;
+               pm_signal_local[j].sub_unit = pm_signal[j].sub_unit;
+               pm_signal_local[j].bit = pm_signal[j].bit;
+       }
+
+       ret = rtas_ibm_cbe_perftools(SUBFUNC_ACTIVATE, PASSTHRU_ENABLE,
+                                    pm_signal_local,
+                                    count * sizeof(struct pm_signal));
+
+       if (ret)
+               printk(KERN_WARNING "%s: rtas returned: %d\n",
+                      __FUNCTION__, ret);
+}
+
+/*
+ * PM Signal functions
+ */
+static void set_pm_event(u32 ctr, int event, u32 unit_mask)
+{
+       struct pm_signal *p;
+       u32 signal_bit;
+       u32 bus_word, bus_type, count_cycles, polarity, input_control;
+       int j, i;
+
+       if (event == PPU_CYCLES_EVENT_NUM) {
+               /* Special Event: Count all cpu cycles */
+               pm_regs.pm07_cntrl[ctr] = CBE_COUNT_ALL_CYCLES;
+               p = &(pm_signal[ctr]);
+               p->signal_group = 21;
+               p->bus_word = 1;
+               p->sub_unit = 0;
+               p->bit = 0;
+               goto out;
+       } else {
+               pm_regs.pm07_cntrl[ctr] = 0;
+       }
+
+       bus_word = GET_BUS_WORD(unit_mask);
+       bus_type = GET_BUS_TYPE(unit_mask);
+       count_cycles = GET_COUNT_CYCLES(unit_mask);
+       polarity = GET_POLARITY(unit_mask);
+       input_control = GET_INPUT_CONTROL(unit_mask);
+       signal_bit = (event % 100);
+
+       p = &(pm_signal[ctr]);
+
+       p->signal_group = event / 100;
+       p->bus_word = bus_word;
+       p->sub_unit = unit_mask & 0x0000f000;
+
+       pm_regs.pm07_cntrl[ctr] = 0;
+       pm_regs.pm07_cntrl[ctr] |= PM07_CTR_COUNT_CYCLES(count_cycles);
+       pm_regs.pm07_cntrl[ctr] |= PM07_CTR_POLARITY(polarity);
+       pm_regs.pm07_cntrl[ctr] |= PM07_CTR_INPUT_CONTROL(input_control);
+
+       if (input_control == 0) {
+               if (signal_bit > 31) {
+                       signal_bit -= 32;
+                       if (bus_word == 0x3)
+                               bus_word = 0x2;
+                       else if (bus_word == 0xc)
+                               bus_word = 0x8;
+               }
+
+               if ((bus_type == 0) && p->signal_group >= 60)
+                       bus_type = 2;
+               if ((bus_type == 1) && p->signal_group >= 50)
+                       bus_type = 0;
+
+               pm_regs.pm07_cntrl[ctr] |= PM07_CTR_INPUT_MUX(signal_bit);
+       } else {
+               pm_regs.pm07_cntrl[ctr] = 0;
+               p->bit = signal_bit;
+       }
+
+       for (i = 0; i < 4; i++) {
+               if (bus_word & (1 << i)) {
+                       pm_regs.debug_bus_control |=
+                           (bus_type << (31 - (2 * i) + 1));
+
+                       for (j = 0; j < 2; j++) {
+                               if (input_bus[j] == 0xff) {
+                                       input_bus[j] = i;
+                                       pm_regs.group_control |=
+                                           (i << (31 - i));
+                                       break;
+                               }
+                       }
+               }
+       }
+out:
+       ;
+}
+
+static void write_pm_cntrl(int cpu, struct pm_cntrl *pm_cntrl)
+{
+       /* Oprofile will use 32 bit counters, set bits 7:10 to 0 */
+       u32 val = 0;
+       if (pm_cntrl->enable == 1)
+               val |= CBE_PM_ENABLE_PERF_MON;
+
+       if (pm_cntrl->stop_at_max == 1)
+               val |= CBE_PM_STOP_AT_MAX;
+
+       if (pm_cntrl->trace_mode == 1)
+               val |= CBE_PM_TRACE_MODE_SET(pm_cntrl->trace_mode);
+
+       if (pm_cntrl->freeze == 1)
+               val |= CBE_PM_FREEZE_ALL_CTRS;
+
+       /* Routine set_count_mode must be called previously to set
+        * the count mode based on the user selection of user and kernel.
+        */
+       val |= CBE_PM_COUNT_MODE_SET(pm_cntrl->count_mode);
+       cbe_write_pm(cpu, pm_control, val);
+}
+
+static inline void
+set_count_mode(u32 kernel, u32 user, struct pm_cntrl *pm_cntrl)
+{
+       /* The user must specify user and kernel if they want them. If
+        *  neither is specified, OProfile will count in hypervisor mode
+        */
+       if (kernel) {
+               if (user)
+                       pm_cntrl->count_mode = CBE_COUNT_ALL_MODES;
+               else
+                       pm_cntrl->count_mode = CBE_COUNT_SUPERVISOR_MODE;
+       } else {
+               if (user)
+                       pm_cntrl->count_mode = CBE_COUNT_PROBLEM_MODE;
+               else
+                       pm_cntrl->count_mode = CBE_COUNT_HYPERVISOR_MODE;
+       }
+}
+
+static inline void enable_ctr(u32 cpu, u32 ctr, u32 * pm07_cntrl)
+{
+
+       pm07_cntrl[ctr] |= PM07_CTR_ENABLE(1);
+       cbe_write_pm07_control(cpu, ctr, pm07_cntrl[ctr]);
+}
+
+/*
+ * Oprofile is expected to collect data on all CPUs simultaneously.
+ * However, there is one set of performance counters per node.  There are
+ * two hardware threads or virtual CPUs on each node.  Hence, OProfile must
+ * multiplex in time the performance counter collection on the two virtual
+ * CPUs.  The multiplexing of the performance counters is done by this
+ * virtual counter routine.
+ *
+ * The pmc_values used below is defined as 'per-cpu' but its use is
+ * more akin to 'per-node'.  We need to store two sets of counter
+ * values per node -- one for the previous run and one for the next.
+ * The per-cpu[NR_PHYS_CTRS] gives us the storage we need.  Each odd/even
+ * pair of per-cpu arrays is used for storing the previous and next
+ * pmc values for a given node.
+ * NOTE: We use the per-cpu variable to improve cache performance.
+ */
+static void cell_virtual_cntr(unsigned long data)
+{
+       /* This routine will alternate loading the virtual counters for
+        * virtual CPUs
+        */
+       int i, prev_hdw_thread, next_hdw_thread;
+       u32 cpu;
+       unsigned long flags;
+
+       /* Make sure that the interrupt_hander and
+        * the virt counter are not both playing with
+        * the counters on the same node.
+        */
+
+       spin_lock_irqsave(&virt_cntr_lock, flags);
+
+       prev_hdw_thread = hdw_thread;
+
+       /* switch the cpu handling the interrupts */
+       hdw_thread = 1 ^ hdw_thread;
+       next_hdw_thread = hdw_thread;
+
+       /* The following is done only once per each node, but
+        * we need cpu #, not node #, to pass to the cbe_xxx functions.
+        */
+       for_each_online_cpu(cpu) {
+               if (cbe_get_hw_thread_id(cpu))
+                       continue;
+
+               /* stop counters, save counter values, restore counts
+                * for previous thread
+                */
+               cbe_disable_pm(cpu);
+               cbe_disable_pm_interrupts(cpu);
+               for (i = 0; i < num_counters; i++) {
+                       per_cpu(pmc_values, cpu + prev_hdw_thread)[i]
+                           = cbe_read_ctr(cpu, i);
+
+                       if (per_cpu(pmc_values, cpu + next_hdw_thread)[i]
+                           == 0xFFFFFFFF)
+                               /* If the cntr value is 0xffffffff, we must
+                                * reset that to 0xfffffff0 when the current
+                                * thread is restarted.  This will generate a new
+                                * interrupt and make sure that we never restore
+                                * the counters to the max value.  If the counters
+                                * were restored to the max value, they do not
+                                * increment and no interrupts are generated.  Hence
+                                * no more samples will be collected on that cpu.
+                                */
+                               cbe_write_ctr(cpu, i, 0xFFFFFFF0);
+                       else
+                               cbe_write_ctr(cpu, i,
+                                             per_cpu(pmc_values,
+                                                     cpu +
+                                                     next_hdw_thread)[i]);
+               }
+
+               /* Switch to the other thread. Change the interrupt
+                * and control regs to be scheduled on the CPU
+                * corresponding to the thread to execute.
+                */
+               for (i = 0; i < num_counters; i++) {
+                       if (pmc_cntrl[next_hdw_thread][i].enabled) {
+                               /* There are some per thread events.
+                                * Must do the set event, enable_cntr
+                                * for each cpu.
+                                */
+                               set_pm_event(i,
+                                    pmc_cntrl[next_hdw_thread][i].evnts,
+                                    pmc_cntrl[next_hdw_thread][i].masks);
+                               enable_ctr(cpu, i,
+                                          pm_regs.pm07_cntrl);
+                       } else {
+                               cbe_write_pm07_control(cpu, i, 0);
+                       }
+               }
+
+               /* Enable interrupts on the CPU thread that is starting */
+               cbe_enable_pm_interrupts(cpu, next_hdw_thread,
+                                        virt_cntr_inter_mask);
+               cbe_enable_pm(cpu);
+       }
+
+       spin_unlock_irqrestore(&virt_cntr_lock, flags);
+
+       mod_timer(&timer_virt_cntr, jiffies + HZ / 10);
+}
+
+static void start_virt_cntrs(void)
+{
+       init_timer(&timer_virt_cntr);
+       timer_virt_cntr.function = cell_virtual_cntr;
+       timer_virt_cntr.data = 0UL;
+       timer_virt_cntr.expires = jiffies + HZ / 10;
+       add_timer(&timer_virt_cntr);
+}
+
+/* This function is called once for all cpus combined */
+static void
+cell_reg_setup(struct op_counter_config *ctr,
+              struct op_system_config *sys, int num_ctrs)
+{
+       int i, j, cpu;
+
+       pm_rtas_token = rtas_token("ibm,cbe-perftools");
+       if (pm_rtas_token == RTAS_UNKNOWN_SERVICE) {
+               printk(KERN_WARNING "%s: RTAS_UNKNOWN_SERVICE\n",
+                      __FUNCTION__);
+               goto out;
+       }
+
+       num_counters = num_ctrs;
+
+       pm_regs.group_control = 0;
+       pm_regs.debug_bus_control = 0;
+
+       /* setup the pm_control register */
+       memset(&pm_regs.pm_cntrl, 0, sizeof(struct pm_cntrl));
+       pm_regs.pm_cntrl.stop_at_max = 1;
+       pm_regs.pm_cntrl.trace_mode = 0;
+       pm_regs.pm_cntrl.freeze = 1;
+
+       set_count_mode(sys->enable_kernel, sys->enable_user,
+                      &pm_regs.pm_cntrl);
+
+       /* Setup the thread 0 events */
+       for (i = 0; i < num_ctrs; ++i) {
+
+               pmc_cntrl[0][i].evnts = ctr[i].event;
+               pmc_cntrl[0][i].masks = ctr[i].unit_mask;
+               pmc_cntrl[0][i].enabled = ctr[i].enabled;
+               pmc_cntrl[0][i].vcntr = i;
+
+               for_each_possible_cpu(j)
+                       per_cpu(pmc_values, j)[i] = 0;
+       }
+
+       /* Setup the thread 1 events, map the thread 0 event to the
+        * equivalent thread 1 event.
+        */
+       for (i = 0; i < num_ctrs; ++i) {
+               if ((ctr[i].event >= 2100) && (ctr[i].event <= 2111))
+                       pmc_cntrl[1][i].evnts = ctr[i].event + 19;
+               else if (ctr[i].event == 2203)
+                       pmc_cntrl[1][i].evnts = ctr[i].event;
+               else if ((ctr[i].event >= 2200) && (ctr[i].event <= 2215))
+                       pmc_cntrl[1][i].evnts = ctr[i].event + 16;
+               else
+                       pmc_cntrl[1][i].evnts = ctr[i].event;
+
+               pmc_cntrl[1][i].masks = ctr[i].unit_mask;
+               pmc_cntrl[1][i].enabled = ctr[i].enabled;
+               pmc_cntrl[1][i].vcntr = i;
+       }
+
+       for (i = 0; i < 4; i++)
+               trace_bus[i] = 0xff;
+
+       for (i = 0; i < 2; i++)
+               input_bus[i] = 0xff;
+
+       /* Our counters count up, and "count" refers to
+        * how much before the next interrupt, and we interrupt
+        * on overflow.  So we calculate the starting value
+        * which will give us "count" until overflow.
+        * Then we set the events on the enabled counters.
+        */
+       for (i = 0; i < num_counters; ++i) {
+               /* start with virtual counter set 0 */
+               if (pmc_cntrl[0][i].enabled) {
+                       /* Using 32bit counters, reset max - count */
+                       reset_value[i] = 0xFFFFFFFF - ctr[i].count;
+                       set_pm_event(i,
+                                    pmc_cntrl[0][i].evnts,
+                                    pmc_cntrl[0][i].masks);
+
+                       /* global, used by cell_cpu_setup */
+                       ctr_enabled |= (1 << i);
+               }
+       }
+
+       /* initialize the previous counts for the virtual cntrs */
+       for_each_online_cpu(cpu)
+               for (i = 0; i < num_counters; ++i) {
+                       per_cpu(pmc_values, cpu)[i] = reset_value[i];
+               }
+out:
+       ;
+}
+
+/* This function is called once for each cpu */
+static void cell_cpu_setup(struct op_counter_config *cntr)
+{
+       u32 cpu = smp_processor_id();
+       u32 num_enabled = 0;
+       int i;
+
+       /* There is one performance monitor per processor chip (i.e. node),
+        * so we only need to perform this function once per node.
+        */
+       if (cbe_get_hw_thread_id(cpu))
+               goto out;
+
+       if (pm_rtas_token == RTAS_UNKNOWN_SERVICE) {
+               printk(KERN_WARNING "%s: RTAS_UNKNOWN_SERVICE\n",
+                      __FUNCTION__);
+               goto out;
+       }
+
+       /* Stop all counters */
+       cbe_disable_pm(cpu);
+       cbe_disable_pm_interrupts(cpu);
+
+       cbe_write_pm(cpu, pm_interval, 0);
+       cbe_write_pm(cpu, pm_start_stop, 0);
+       cbe_write_pm(cpu, group_control, pm_regs.group_control);
+       cbe_write_pm(cpu, debug_bus_control, pm_regs.debug_bus_control);
+       write_pm_cntrl(cpu, &pm_regs.pm_cntrl);
+
+       for (i = 0; i < num_counters; ++i) {
+               if (ctr_enabled & (1 << i)) {
+                       pm_signal[num_enabled].cpu = cbe_cpu_to_node(cpu);
+                       num_enabled++;
+               }
+       }
+
+       pm_rtas_activate_signals(cbe_cpu_to_node(cpu), num_enabled);
+out:
+       ;
+}
+
+static void cell_global_start(struct op_counter_config *ctr)
+{
+       u32 cpu;
+       u32 interrupt_mask = 0;
+       u32 i;
+
+       /* This routine gets called once for the system.
+        * There is one performance monitor per node, so we
+        * only need to perform this function once per node.
+        */
+       for_each_online_cpu(cpu) {
+               if (cbe_get_hw_thread_id(cpu))
+                       continue;
+
+               interrupt_mask = 0;
+
+               for (i = 0; i < num_counters; ++i) {
+                       if (ctr_enabled & (1 << i)) {
+                               cbe_write_ctr(cpu, i, reset_value[i]);
+                               enable_ctr(cpu, i, pm_regs.pm07_cntrl);
+                               interrupt_mask |=
+                                   CBE_PM_CTR_OVERFLOW_INTR(i);
+                       } else {
+                               /* Disable counter */
+                               cbe_write_pm07_control(cpu, i, 0);
+                       }
+               }
+
+               cbe_clear_pm_interrupts(cpu);
+               cbe_enable_pm_interrupts(cpu, hdw_thread, interrupt_mask);
+               cbe_enable_pm(cpu);
+       }
+
+       virt_cntr_inter_mask = interrupt_mask;
+       oprofile_running = 1;
+       smp_wmb();
+
+       /* NOTE: start_virt_cntrs will result in cell_virtual_cntr() being
+        * executed which manipulates the PMU.  We start the "virtual counter"
+        * here so that we do not need to synchronize access to the PMU in
+        * the above for-loop.
+        */
+       start_virt_cntrs();
+}
+
+static void cell_global_stop(void)
+{
+       int cpu;
+
+       /* This routine will be called once for the system.
+        * There is one performance monitor per node, so we
+        * only need to perform this function once per node.
+        */
+       del_timer_sync(&timer_virt_cntr);
+       oprofile_running = 0;
+       smp_wmb();
+
+       for_each_online_cpu(cpu) {
+               if (cbe_get_hw_thread_id(cpu))
+                       continue;
+
+               cbe_sync_irq(cbe_cpu_to_node(cpu));
+               /* Stop the counters */
+               cbe_disable_pm(cpu);
+
+               /* Deactivate the signals */
+               pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+
+               /* Deactivate interrupts */
+               cbe_disable_pm_interrupts(cpu);
+       }
+}
+
+static void
+cell_handle_interrupt(struct pt_regs *regs, struct op_counter_config *ctr)
+{
+       u32 cpu;
+       u64 pc;
+       int is_kernel;
+       unsigned long flags = 0;
+       u32 interrupt_mask;
+       int i;
+
+       cpu = smp_processor_id();
+
+       /* Need to make sure the interrupt handler and the virt counter
+        * routine are not running at the same time. See the
+        * cell_virtual_cntr() routine for additional comments.
+        */
+       spin_lock_irqsave(&virt_cntr_lock, flags);
+
+       /* Need to disable and reenable the performance counters
+        * to get the desired behavior from the hardware.  This
+        * is hardware specific.
+        */
+
+       cbe_disable_pm(cpu);
+
+       interrupt_mask = cbe_clear_pm_interrupts(cpu);
+
+       /* If the interrupt mask has been cleared, then the virt cntr
+        * has cleared the interrupt.  When the thread that generated
+        * the interrupt is restored, the data count will be restored to
+        * 0xffffff0 to cause the interrupt to be regenerated.
+        */
+
+       if ((oprofile_running == 1) && (interrupt_mask != 0)) {
+               pc = regs->nip;
+               is_kernel = is_kernel_addr(pc);
+
+               for (i = 0; i < num_counters; ++i) {
+                       if ((interrupt_mask & CBE_PM_CTR_OVERFLOW_INTR(i))
+                           && ctr[i].enabled) {
+                               oprofile_add_pc(pc, is_kernel, i);
+                               cbe_write_ctr(cpu, i, reset_value[i]);
+                       }
+               }
+
+               /* The counters were frozen by the interrupt.
+                * Reenable the interrupt and restart the counters.
+                * If there was a race between the interrupt handler and
+                * the virtual counter routine.  The virutal counter
+                * routine may have cleared the interrupts.  Hence must
+                * use the virt_cntr_inter_mask to re-enable the interrupts.
+                */
+               cbe_enable_pm_interrupts(cpu, hdw_thread,
+                                        virt_cntr_inter_mask);
+
+               /* The writes to the various performance counters only writes
+                * to a latch.  The new values (interrupt setting bits, reset
+                * counter value etc.) are not copied to the actual registers
+                * until the performance monitor is enabled.  In order to get
+                * this to work as desired, the permormance monitor needs to
+                * be disabled while writting to the latches.  This is a
+                * HW design issue.
+                */
+               cbe_enable_pm(cpu);
+       }
+       spin_unlock_irqrestore(&virt_cntr_lock, flags);
+}
+
+struct op_powerpc_model op_model_cell = {
+       .reg_setup = cell_reg_setup,
+       .cpu_setup = cell_cpu_setup,
+       .global_start = cell_global_start,
+       .global_stop = cell_global_stop,
+       .handle_interrupt = cell_handle_interrupt,
+};
index 5a91b75c2f0125f84c914cc3e7092c3c8ee5a32f..9a0ee62691d529f69a90779e5bbe62cd7e5fbdcd 100644 (file)
@@ -130,6 +130,18 @@ struct cbe_mic_tm_regs __iomem *cbe_get_cpu_mic_tm_regs(int cpu)
 }
 EXPORT_SYMBOL_GPL(cbe_get_cpu_mic_tm_regs);
 
+/* FIXME
+ * This is little more than a stub at the moment.  It should be
+ * fleshed out so that it works for both SMT and non-SMT, no
+ * matter if the passed cpu is odd or even.
+ * For SMT enabled, returns 0 for even-numbered cpu; otherwise 1.
+ * For SMT disabled, returns 0 for all cpus.
+ */
+u32 cbe_get_hw_thread_id(int cpu)
+{
+       return (cpu & 1);
+}
+EXPORT_SYMBOL_GPL(cbe_get_hw_thread_id);
 
 void __init cbe_regs_init(void)
 {
index f28abf2fc273a7599277c01a82ade0703262ceee..99c612025e8fa824dc35cb34741fc6672f6a2e39 100644 (file)
@@ -25,6 +25,7 @@
 #include <linux/interrupt.h>
 #include <linux/types.h>
 #include <asm/io.h>
+#include <asm/irq_regs.h>
 #include <asm/machdep.h>
 #include <asm/pmc.h>
 #include <asm/reg.h>
@@ -375,9 +376,9 @@ void cbe_disable_pm_interrupts(u32 cpu)
 }
 EXPORT_SYMBOL_GPL(cbe_disable_pm_interrupts);
 
-static irqreturn_t cbe_pm_irq(int irq, void *dev_id, struct pt_regs *regs)
+static irqreturn_t cbe_pm_irq(int irq, void *dev_id)
 {
-       perf_irq(regs);
+       perf_irq(get_irq_regs());
        return IRQ_HANDLED;
 }
 
@@ -408,3 +409,21 @@ int __init cbe_init_pm_irq(void)
 }
 arch_initcall(cbe_init_pm_irq);
 
+void cbe_sync_irq(int node)
+{
+       unsigned int irq;
+
+       irq = irq_find_mapping(NULL,
+                              IIC_IRQ_IOEX_PMI
+                              | (node << IIC_IRQ_NODE_SHIFT));
+
+       if (irq == NO_IRQ) {
+               printk(KERN_WARNING "ERROR, unable to get existing irq %d " \
+               "for node %d\n", irq, node);
+               return;
+       }
+
+       synchronize_irq(irq);
+}
+EXPORT_SYMBOL_GPL(cbe_sync_irq);
+
index 63037c352fde8bd7a2e0d1a56d4dc7e6863e9161..e8c2ebd3dddac8998c92547c7f08e18302af6fa0 100644 (file)
@@ -91,5 +91,23 @@ extern void cbe_enable_pm_interrupts(u32 cpu, u32 thread, u32 mask);
 extern void cbe_disable_pm_interrupts(u32 cpu);
 extern u32  cbe_query_pm_interrupts(u32 cpu);
 extern u32  cbe_clear_pm_interrupts(u32 cpu);
+extern void cbe_sync_irq(int node);
+
+/* Utility functions, macros */
+extern u32 cbe_get_hw_thread_id(int cpu);
+
+#define cbe_cpu_to_node(cpu) ((cpu) >> 1)
+
+#define CBE_COUNT_SUPERVISOR_MODE       0
+#define CBE_COUNT_HYPERVISOR_MODE       1
+#define CBE_COUNT_PROBLEM_MODE          2
+#define CBE_COUNT_ALL_MODES             3
+
+/* Macros for the pm07_control registers. */
+#define PM07_CTR_INPUT_MUX(x)                    (((x) & 0x3F) << 26)
+#define PM07_CTR_INPUT_CONTROL(x)                (((x) & 1) << 25)
+#define PM07_CTR_POLARITY(x)                     (((x) & 1) << 24)
+#define PM07_CTR_COUNT_CYCLES(x)                 (((x) & 1) << 23)
+#define PM07_CTR_ENABLE(x)                       (((x) & 1) << 22)
 
 #endif /* __ASM_CELL_PMU_H__ */
index a9a40149a7c0daf66cca975035b4d5863bbc10bc..762d15e488791d39464fbcc74c0ccf1d19138f31 100644 (file)
@@ -45,6 +45,7 @@ enum powerpc_oprofile_type {
        PPC_OPROFILE_POWER4 = 2,
        PPC_OPROFILE_G4 = 3,
        PPC_OPROFILE_BOOKE = 4,
+       PPC_OPROFILE_CELL = 5,
 };
 
 struct cpu_spec {
index 07a10e590c1de44d951bb0b25e3bdfd364670f94..71043bf3641f369311c3343c6b0700ff899ea981 100644 (file)
@@ -44,7 +44,9 @@ struct op_powerpc_model {
                           int num_counters);
        void (*cpu_setup) (struct op_counter_config *);
        void (*start) (struct op_counter_config *);
+        void (*global_start) (struct op_counter_config *);
        void (*stop) (void);
+       void (*global_stop) (void);
        void (*handle_interrupt) (struct pt_regs *,
                                  struct op_counter_config *);
        int num_counters;
@@ -54,6 +56,7 @@ extern struct op_powerpc_model op_model_fsl_booke;
 extern struct op_powerpc_model op_model_rs64;
 extern struct op_powerpc_model op_model_power4;
 extern struct op_powerpc_model op_model_7450;
+extern struct op_powerpc_model op_model_cell;
 
 #ifndef CONFIG_FSL_BOOKE