#define SPU_PROFILING_CYCLES 1
#define SPU_PROFILING_EVENTS 2
+#define SPU_EVENT_NUM_START 4100
+#define SPU_EVENT_NUM_STOP 4399
+#define SPU_PROFILE_EVENT_ADDR 4363 /* spu, address trace, decimal */
+#define SPU_PROFILE_EVENT_ADDR_MASK_A 0x146 /* sub unit set to zero */
+#define SPU_PROFILE_EVENT_ADDR_MASK_B 0x186 /* sub unit set to zero */
+
#define NUM_SPUS_PER_NODE 8
#define SPU_CYCLES_EVENT_NUM 2 /* event number for SPU_CYCLES */
#define MAX_SPU_COUNT 0xFFFFFF /* maximum 24 bit LFSR value */
+/* Minumum HW interval timer setting to send value to trace buffer is 10 cycle.
+ * To configure counter to send value every N cycles set counter to
+ * 2^32 - 1 - N.
+ */
+#define NUM_INTERVAL_CYC 0xFFFFFFFF - 10
+
/*
* spu_cycle_reset is the number of cycles between samples.
* This variable is used for SPU profiling and should ONLY be set
*/
static unsigned int spu_cycle_reset;
static unsigned int profiling_mode;
+static int spu_evnt_phys_spu_indx;
struct pmc_cntrl_data {
unsigned long vcntr;
u16 trace_mode;
u16 freeze;
u16 count_mode;
+ u16 spu_addr_trace;
+ u8 trace_buf_ovflw;
};
static struct {
#define GET_INPUT_CONTROL(x) ((x & 0x00000004) >> 2)
static DEFINE_PER_CPU(unsigned long[NR_PHYS_CTRS], pmc_values);
+static unsigned long spu_pm_cnt[MAX_NUMNODES * NUM_SPUS_PER_NODE];
static struct pmc_cntrl_data pmc_cntrl[NUM_THREADS][NR_PHYS_CTRS];
/*
static u32 virt_cntr_inter_mask;
static struct timer_list timer_virt_cntr;
+static struct timer_list timer_spu_event_swap;
/*
* pm_signal needs to be global since it is initialized in
if (pm_regs.pm_cntrl.trace_mode != 0)
val |= CBE_PM_TRACE_MODE_SET(pm_regs.pm_cntrl.trace_mode);
+ if (pm_regs.pm_cntrl.trace_buf_ovflw == 1)
+ val |= CBE_PM_TRACE_BUF_OVFLW(pm_regs.pm_cntrl.trace_buf_ovflw);
if (pm_regs.pm_cntrl.freeze == 1)
val |= CBE_PM_FREEZE_ALL_CTRS;
+ val |= CBE_PM_SPU_ADDR_TRACE_SET(pm_regs.pm_cntrl.spu_addr_trace);
+
/*
* Routine set_count_mode must be called previously to set
* the count mode based on the user selection of user and kernel.
return 0;
}
+/* Unfortunately, the hardware will only support event profiling
+ * on one SPU per node at a time. Therefore, we must time slice
+ * the profiling across all SPUs in the node. Note, we do this
+ * in parallel for each node. The following routine is called
+ * periodically based on kernel timer to switch which SPU is
+ * being monitored in a round robbin fashion.
+ */
+static void spu_evnt_swap(unsigned long data)
+{
+ int node;
+ int cur_phys_spu, nxt_phys_spu, cur_spu_evnt_phys_spu_indx;
+ unsigned long flags;
+ int cpu;
+ int ret;
+ u32 interrupt_mask;
+
+
+ /* enable interrupts on cntr 0 */
+ interrupt_mask = CBE_PM_CTR_OVERFLOW_INTR(0);
+
+ hdw_thread = 0;
+
+ /* Make sure spu event interrupt handler and spu event swap
+ * don't access the counters simultaneously.
+ */
+ spin_lock_irqsave(&cntr_lock, flags);
+
+ cur_spu_evnt_phys_spu_indx = spu_evnt_phys_spu_indx;
+
+ if (++(spu_evnt_phys_spu_indx) == NUM_SPUS_PER_NODE)
+ spu_evnt_phys_spu_indx = 0;
+
+ pm_signal[0].sub_unit = spu_evnt_phys_spu_indx;
+ pm_signal[1].sub_unit = spu_evnt_phys_spu_indx;
+ pm_signal[2].sub_unit = spu_evnt_phys_spu_indx;
+
+ /* switch the SPU being profiled on each node */
+ for_each_online_cpu(cpu) {
+ if (cbe_get_hw_thread_id(cpu))
+ continue;
+
+ node = cbe_cpu_to_node(cpu);
+ cur_phys_spu = (node * NUM_SPUS_PER_NODE)
+ + cur_spu_evnt_phys_spu_indx;
+ nxt_phys_spu = (node * NUM_SPUS_PER_NODE)
+ + spu_evnt_phys_spu_indx;
+
+ /*
+ * stop counters, save counter values, restore counts
+ * for previous physical SPU
+ */
+ cbe_disable_pm(cpu);
+ cbe_disable_pm_interrupts(cpu);
+
+ spu_pm_cnt[cur_phys_spu]
+ = cbe_read_ctr(cpu, 0);
+
+ /* restore previous count for the next spu to sample */
+ /* NOTE, hardware issue, counter will not start if the
+ * counter value is at max (0xFFFFFFFF).
+ */
+ if (spu_pm_cnt[nxt_phys_spu] >= 0xFFFFFFFF)
+ cbe_write_ctr(cpu, 0, 0xFFFFFFF0);
+ else
+ cbe_write_ctr(cpu, 0, spu_pm_cnt[nxt_phys_spu]);
+
+ pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+
+ /* setup the debug bus measure the one event and
+ * the two events to route the next SPU's PC on
+ * the debug bus
+ */
+ ret = pm_rtas_activate_signals(cbe_cpu_to_node(cpu), 3);
+ if (ret)
+ printk(KERN_ERR
+ "%s: pm_rtas_activate_signals failed, SPU event swap\n",
+ __func__);
+
+ /* clear the trace buffer, don't want to take PC for
+ * previous SPU*/
+ cbe_write_pm(cpu, trace_address, 0);
+
+ enable_ctr(cpu, 0, pm_regs.pm07_cntrl);
+
+ /* Enable interrupts on the CPU thread that is starting */
+ cbe_enable_pm_interrupts(cpu, hdw_thread,
+ interrupt_mask);
+ cbe_enable_pm(cpu);
+ }
+
+ spin_unlock_irqrestore(&cntr_lock, flags);
+
+ /* swap approximately every 0.1 seconds */
+ mod_timer(&timer_spu_event_swap, jiffies + HZ / 25);
+}
+
+static void start_spu_event_swap(void)
+{
+ init_timer(&timer_spu_event_swap);
+ timer_spu_event_swap.function = spu_evnt_swap;
+ timer_spu_event_swap.data = 0UL;
+ timer_spu_event_swap.expires = jiffies + HZ / 25;
+ add_timer(&timer_spu_event_swap);
+}
+
+static int cell_reg_setup_spu_events(struct op_counter_config *ctr,
+ struct op_system_config *sys, int num_ctrs)
+{
+ int i;
+
+ /* routine is called once for all nodes */
+
+ spu_evnt_phys_spu_indx = 0;
+ /*
+ * For all events except PPU CYCLEs, each node will need to make
+ * the rtas cbe-perftools call to setup and reset the debug bus.
+ * Make the token lookup call once and store it in the global
+ * variable pm_rtas_token.
+ */
+ pm_rtas_token = rtas_token("ibm,cbe-perftools");
+
+ if (unlikely(pm_rtas_token == RTAS_UNKNOWN_SERVICE)) {
+ printk(KERN_ERR
+ "%s: rtas token ibm,cbe-perftools unknown\n",
+ __func__);
+ return -EIO;
+ }
+
+ /* setup the pm_control register settings,
+ * settings will be written per node by the
+ * cell_cpu_setup() function.
+ */
+ pm_regs.pm_cntrl.trace_buf_ovflw = 1;
+
+ /* Use the occurrence trace mode to have SPU PC saved
+ * to the trace buffer. Occurrence data in trace buffer
+ * is not used. Bit 2 must be set to store SPU addresses.
+ */
+ pm_regs.pm_cntrl.trace_mode = 2;
+
+ pm_regs.pm_cntrl.spu_addr_trace = 0x1; /* using debug bus
+ event 2 & 3 */
+
+ /* setup the debug bus event array with the SPU PC routing events.
+ * Note, pm_signal[0] will be filled in by set_pm_event() call below.
+ */
+ pm_signal[1].signal_group = SPU_PROFILE_EVENT_ADDR / 100;
+ pm_signal[1].bus_word = GET_BUS_WORD(SPU_PROFILE_EVENT_ADDR_MASK_A);
+ pm_signal[1].bit = SPU_PROFILE_EVENT_ADDR % 100;
+ pm_signal[1].sub_unit = spu_evnt_phys_spu_indx;
+
+ pm_signal[2].signal_group = SPU_PROFILE_EVENT_ADDR / 100;
+ pm_signal[2].bus_word = GET_BUS_WORD(SPU_PROFILE_EVENT_ADDR_MASK_B);
+ pm_signal[2].bit = SPU_PROFILE_EVENT_ADDR % 100;
+ pm_signal[2].sub_unit = spu_evnt_phys_spu_indx;
+
+ /* Set the user selected spu event to profile on,
+ * note, only one SPU profiling event is supported
+ */
+ num_counters = 1; /* Only support one SPU event at a time */
+ set_pm_event(0, ctr[0].event, ctr[0].unit_mask);
+
+ reset_value[0] = 0xFFFFFFFF - ctr[0].count;
+
+ /* global, used by cell_cpu_setup */
+ ctr_enabled |= 1;
+
+ /* Initialize the count for each SPU to the reset value */
+ for (i=0; i < MAX_NUMNODES * NUM_SPUS_PER_NODE; i++)
+ spu_pm_cnt[i] = reset_value[0];
+
+ return 0;
+}
+
static int cell_reg_setup_ppu(struct op_counter_config *ctr,
struct op_system_config *sys, int num_ctrs)
{
+ /* routine is called once for all nodes */
int i, j, cpu;
num_counters = num_ctrs;
__func__);
return -EIO;
}
- 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);
static int cell_reg_setup(struct op_counter_config *ctr,
struct op_system_config *sys, int num_ctrs)
{
- int ret;
-
+ int ret=0;
spu_cycle_reset = 0;
+ /* initialize the spu_arr_trace value, will be reset if
+ * doing spu event profiling.
+ */
+ pm_regs.group_control = 0;
+ pm_regs.debug_bus_control = 0;
+ pm_regs.pm_cntrl.stop_at_max = 1;
+ pm_regs.pm_cntrl.trace_mode = 0;
+ pm_regs.pm_cntrl.freeze = 1;
+ pm_regs.pm_cntrl.trace_buf_ovflw = 0;
+ pm_regs.pm_cntrl.spu_addr_trace = 0;
+
/*
* For all events except PPU CYCLEs, each node will need to make
* the rtas cbe-perftools call to setup and reset the debug bus.
if (ctr[0].event == SPU_CYCLES_EVENT_NUM) {
profiling_mode = SPU_PROFILING_CYCLES;
ret = cell_reg_setup_spu_cycles(ctr, sys, num_ctrs);
+ } else if ((ctr[0].event >= SPU_EVENT_NUM_START) &&
+ (ctr[0].event <= SPU_EVENT_NUM_STOP)) {
+ profiling_mode = SPU_PROFILING_EVENTS;
+ spu_cycle_reset = ctr[0].count;
+
+ /* for SPU event profiling, need to setup the
+ * pm_signal array with the events to route the
+ * SPU PC before making the FW call. Note, only
+ * one SPU event for profiling can be specified
+ * at a time.
+ */
+ cell_reg_setup_spu_events(ctr, sys, num_ctrs);
} else {
profiling_mode = PPU_PROFILING;
ret = cell_reg_setup_ppu(ctr, sys, num_ctrs);
u32 cpu = smp_processor_id();
u32 num_enabled = 0;
int i;
+ int ret;
/* Cycle based SPU profiling does not use the performance
* counters. The trace array is configured to collect
* The pm_rtas_activate_signals will return -EIO if the FW
* call failed.
*/
- return pm_rtas_activate_signals(cbe_cpu_to_node(cpu), num_enabled);
+ if (profiling_mode == SPU_PROFILING_EVENTS) {
+ /* For SPU event profiling also need to setup the
+ * pm interval timer
+ */
+ ret = pm_rtas_activate_signals(cbe_cpu_to_node(cpu),
+ num_enabled+2);
+ /* store PC from debug bus to Trace buffer as often
+ * as possible (every 10 cycles)
+ */
+ cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC);
+ return ret;
+ } else
+ return pm_rtas_activate_signals(cbe_cpu_to_node(cpu),
+ num_enabled);
}
#define ENTRIES 303
int cpu;
oprofile_running = 0;
+ smp_wmb();
#ifdef CONFIG_CPU_FREQ
cpufreq_unregister_notifier(&cpu_freq_notifier_block,
pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
}
- if (profiling_mode == SPU_PROFILING_CYCLES)
- stop_spu_profiling_cycles();
+ stop_spu_profiling_cycles();
+}
+
+static void cell_global_stop_spu_events(void)
+{
+ int cpu;
+ oprofile_running = 0;
+
+ stop_spu_profiling_events();
+ 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);
+ cbe_write_pm07_control(cpu, 0, 0);
+
+ /* Deactivate the signals */
+ pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+
+ /* Deactivate interrupts */
+ cbe_disable_pm_interrupts(cpu);
+ }
+ del_timer_sync(&timer_spu_event_swap);
}
static void cell_global_stop_ppu(void)
{
if (profiling_mode == PPU_PROFILING)
cell_global_stop_ppu();
+ else if (profiling_mode == SPU_PROFILING_EVENTS)
+ cell_global_stop_spu_events();
else
cell_global_stop_spu_cycles();
}
return rtas_error;
}
+static int cell_global_start_spu_events(struct op_counter_config *ctr)
+{
+ int cpu;
+ u32 interrupt_mask = 0;
+ int rtn = 0;
+
+ hdw_thread = 0;
+
+ /* spu event profiling, uses the performance counters to generate
+ * an interrupt. The hardware is setup to store the SPU program
+ * counter into the trace array. The occurrence mode is used to
+ * enable storing data to the trace buffer. The bits are set
+ * to send/store the SPU address in the trace buffer. The debug
+ * bus must be setup to route the SPU program counter onto the
+ * debug bus. The occurrence data in the trace buffer is not used.
+ */
+
+ /* 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;
+
+ /*
+ * Setup SPU event-based profiling.
+ * Set perf_mon_control bit 0 to a zero before
+ * enabling spu collection hardware.
+ *
+ * Only support one SPU event on one SPU per node.
+ */
+ if (ctr_enabled & 1) {
+ cbe_write_ctr(cpu, 0, reset_value[0]);
+ enable_ctr(cpu, 0, pm_regs.pm07_cntrl);
+ interrupt_mask |=
+ CBE_PM_CTR_OVERFLOW_INTR(0);
+ } else {
+ /* Disable counter */
+ cbe_write_pm07_control(cpu, 0, 0);
+ }
+
+ cbe_get_and_clear_pm_interrupts(cpu);
+ cbe_enable_pm_interrupts(cpu, hdw_thread, interrupt_mask);
+ cbe_enable_pm(cpu);
+
+ /* clear the trace buffer */
+ cbe_write_pm(cpu, trace_address, 0);
+ }
+
+ /* Start the timer to time slice collecting the event profile
+ * on each of the SPUs. Note, can collect profile on one SPU
+ * per node at a time.
+ */
+ start_spu_event_swap();
+ start_spu_profiling_events();
+ oprofile_running = 1;
+ smp_wmb();
+
+ return rtn;
+}
+
static int cell_global_start_ppu(struct op_counter_config *ctr)
{
u32 cpu, i;
{
if (profiling_mode == SPU_PROFILING_CYCLES)
return cell_global_start_spu_cycles(ctr);
+ else if (profiling_mode == SPU_PROFILING_EVENTS)
+ return cell_global_start_spu_events(ctr);
else
return cell_global_start_ppu(ctr);
}
+/* The SPU interrupt handler
+ *
+ * SPU event profiling works as follows:
+ * The pm_signal[0] holds the one SPU event to be measured. It is routed on
+ * the debug bus using word 0 or 1. The value of pm_signal[1] and
+ * pm_signal[2] contain the necessary events to route the SPU program
+ * counter for the selected SPU onto the debug bus using words 2 and 3.
+ * The pm_interval register is setup to write the SPU PC value into the
+ * trace buffer at the maximum rate possible. The trace buffer is configured
+ * to store the PCs, wrapping when it is full. The performance counter is
+ * intialized to the max hardware count minus the number of events, N, between
+ * samples. Once the N events have occured, a HW counter overflow occurs
+ * causing the generation of a HW counter interrupt which also stops the
+ * writing of the SPU PC values to the trace buffer. Hence the last PC
+ * written to the trace buffer is the SPU PC that we want. Unfortunately,
+ * we have to read from the beginning of the trace buffer to get to the
+ * last value written. We just hope the PPU has nothing better to do then
+ * service this interrupt. The PC for the specific SPU being profiled is
+ * extracted from the trace buffer processed and stored. The trace buffer
+ * is cleared, interrupts are cleared, the counter is reset to max - N.
+ * A kernel timer is used to periodically call the routine spu_evnt_swap()
+ * to switch to the next physical SPU in the node to profile in round robbin
+ * order. This way data is collected for all SPUs on the node. It does mean
+ * that we need to use a relatively small value of N to ensure enough samples
+ * on each SPU are collected each SPU is being profiled 1/8 of the time.
+ * It may also be necessary to use a longer sample collection period.
+ */
+static void cell_handle_interrupt_spu(struct pt_regs *regs,
+ struct op_counter_config *ctr)
+{
+ u32 cpu, cpu_tmp;
+ u64 trace_entry;
+ u32 interrupt_mask;
+ u64 trace_buffer[2];
+ u64 last_trace_buffer;
+ u32 sample;
+ u32 trace_addr;
+ unsigned long sample_array_lock_flags;
+ int spu_num;
+ unsigned long flags;
+
+ /* Make sure spu event interrupt handler and spu event swap
+ * don't access the counters simultaneously.
+ */
+ cpu = smp_processor_id();
+ spin_lock_irqsave(&cntr_lock, flags);
+
+ cpu_tmp = cpu;
+ cbe_disable_pm(cpu);
+
+ interrupt_mask = cbe_get_and_clear_pm_interrupts(cpu);
+
+ sample = 0xABCDEF;
+ trace_entry = 0xfedcba;
+ last_trace_buffer = 0xdeadbeaf;
+
+ if ((oprofile_running == 1) && (interrupt_mask != 0)) {
+ /* disable writes to trace buff */
+ cbe_write_pm(cpu, pm_interval, 0);
+
+ /* only have one perf cntr being used, cntr 0 */
+ if ((interrupt_mask & CBE_PM_CTR_OVERFLOW_INTR(0))
+ && ctr[0].enabled)
+ /* The SPU PC values will be read
+ * from the trace buffer, reset counter
+ */
+
+ cbe_write_ctr(cpu, 0, reset_value[0]);
+
+ trace_addr = cbe_read_pm(cpu, trace_address);
+
+ while (!(trace_addr & CBE_PM_TRACE_BUF_EMPTY)) {
+ /* There is data in the trace buffer to process
+ * Read the buffer until you get to the last
+ * entry. This is the value we want.
+ */
+
+ cbe_read_trace_buffer(cpu, trace_buffer);
+ trace_addr = cbe_read_pm(cpu, trace_address);
+ }
+
+ /* SPU Address 16 bit count format for 128 bit
+ * HW trace buffer is used for the SPU PC storage
+ * HDR bits 0:15
+ * SPU Addr 0 bits 16:31
+ * SPU Addr 1 bits 32:47
+ * unused bits 48:127
+ *
+ * HDR: bit4 = 1 SPU Address 0 valid
+ * HDR: bit5 = 1 SPU Address 1 valid
+ * - unfortunately, the valid bits don't seem to work
+ *
+ * Note trace_buffer[0] holds bits 0:63 of the HW
+ * trace buffer, trace_buffer[1] holds bits 64:127
+ */
+
+ trace_entry = trace_buffer[0]
+ & 0x00000000FFFF0000;
+
+ /* only top 16 of the 18 bit SPU PC address
+ * is stored in trace buffer, hence shift right
+ * by 16 -2 bits */
+ sample = trace_entry >> 14;
+ last_trace_buffer = trace_buffer[0];
+
+ spu_num = spu_evnt_phys_spu_indx
+ + (cbe_cpu_to_node(cpu) * NUM_SPUS_PER_NODE);
+
+ /* make sure only one process at a time is calling
+ * spu_sync_buffer()
+ */
+ spin_lock_irqsave(&oprof_spu_smpl_arry_lck,
+ sample_array_lock_flags);
+ spu_sync_buffer(spu_num, &sample, 1);
+ spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck,
+ sample_array_lock_flags);
+
+ smp_wmb(); /* insure spu event buffer updates are written
+ * don't want events intermingled... */
+
+ /* The counters were frozen by the interrupt.
+ * Reenable the interrupt and restart the counters.
+ */
+ cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC);
+ cbe_enable_pm_interrupts(cpu, hdw_thread,
+ virt_cntr_inter_mask);
+
+ /* clear the trace buffer, re-enable writes to trace buff */
+ cbe_write_pm(cpu, trace_address, 0);
+ cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC);
+
+ /* 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 writing to the latches. This is a
+ * HW design issue.
+ */
+ write_pm_cntrl(cpu);
+ cbe_enable_pm(cpu);
+ }
+ spin_unlock_irqrestore(&cntr_lock, flags);
+}
+
static void cell_handle_interrupt_ppu(struct pt_regs *regs,
struct op_counter_config *ctr)
{
{
if (profiling_mode == PPU_PROFILING)
cell_handle_interrupt_ppu(regs, ctr);
+ else
+ cell_handle_interrupt_spu(regs, ctr);
}
/*
git://ftp.safe.ca / safe / jmp / linux-2.6 / blobdiff