arch/powerpc/oprofile/cell/spu_profiler.c

   1 /*
   2  * Cell Broadband Engine OProfile Support
   3  *
   4  * (C) Copyright IBM Corporation 2006
   5  *
   6  * Authors: Maynard Johnson <maynardj@us.ibm.com>
   7  *          Carl Love <carll@us.ibm.com>
   8  *
   9  * This program is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU General Public License
  11  * as published by the Free Software Foundation; either version
  12  * 2 of the License, or (at your option) any later version.
  13  */
  14
  15 #include <linux/hrtimer.h>
  16 #include <linux/smp.h>
  17 #include <linux/slab.h>
  18 #include <asm/cell-pmu.h>
  19 #include "pr_util.h"
  20
  21 #define TRACE_ARRAY_SIZE 1024
  22 #define SCALE_SHIFT 14
  23
  24 static u32 *samples;
  25
  26 static int spu_prof_running;
  27 static unsigned int profiling_interval;
  28
  29 #define NUM_SPU_BITS_TRBUF 16
  30 #define SPUS_PER_TB_ENTRY   4
  31 #define SPUS_PER_NODE        8
  32
  33 #define SPU_PC_MASK          0xFFFF
  34
  35 static DEFINE_SPINLOCK(sample_array_lock);
  36 unsigned long sample_array_lock_flags;
  37
  38 void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_reset)
  39 {
  40         unsigned long ns_per_cyc;
  41
  42         if (!freq_khz)
  43                 freq_khz = ppc_proc_freq/1000;
  44
  45         /* To calculate a timeout in nanoseconds, the basic
  46          * formula is ns = cycles_reset * (NSEC_PER_SEC / cpu frequency).
  47          * To avoid floating point math, we use the scale math
  48          * technique as described in linux/jiffies.h.  We use
  49          * a scale factor of SCALE_SHIFT, which provides 4 decimal places
  50          * of precision.  This is close enough for the purpose at hand.
  51          *
  52          * The value of the timeout should be small enough that the hw
  53          * trace buffer will not get more then about 1/3 full for the
  54          * maximum user specified (the LFSR value) hw sampling frequency.
  55          * This is to ensure the trace buffer will never fill even if the
  56          * kernel thread scheduling varies under a heavy system load.
  57          */
  58
  59         ns_per_cyc = (USEC_PER_SEC << SCALE_SHIFT)/freq_khz;
  60         profiling_interval = (ns_per_cyc * cycles_reset) >> SCALE_SHIFT;
  61
  62 }
  63
  64 /*
  65  * Extract SPU PC from trace buffer entry
  66  */
  67 static void spu_pc_extract(int cpu, int entry)
  68 {
  69         /* the trace buffer is 128 bits */
  70         u64 trace_buffer[2];
  71         u64 spu_mask;
  72         int spu;
  73
  74         spu_mask = SPU_PC_MASK;
  75
  76         /* Each SPU PC is 16 bits; hence, four spus in each of
  77          * the two 64-bit buffer entries that make up the
  78          * 128-bit trace_buffer entry.  Process two 64-bit values
  79          * simultaneously.
  80          * trace[0] SPU PC contents are: 0 1 2 3
  81          * trace[1] SPU PC contents are: 4 5 6 7
  82          */
  83
  84         cbe_read_trace_buffer(cpu, trace_buffer);
  85
  86         for (spu = SPUS_PER_TB_ENTRY-1; spu >= 0; spu--) {
  87                 /* spu PC trace entry is upper 16 bits of the
  88                  * 18 bit SPU program counter
  89                  */
  90                 samples[spu * TRACE_ARRAY_SIZE + entry]
  91                         = (spu_mask & trace_buffer[0]) << 2;
  92                 samples[(spu + SPUS_PER_TB_ENTRY) * TRACE_ARRAY_SIZE + entry]
  93                         = (spu_mask & trace_buffer[1]) << 2;
  94
  95                 trace_buffer[0] = trace_buffer[0] >> NUM_SPU_BITS_TRBUF;
  96                 trace_buffer[1] = trace_buffer[1] >> NUM_SPU_BITS_TRBUF;
  97         }
  98 }
  99
 100 static int cell_spu_pc_collection(int cpu)
 101 {
 102         u32 trace_addr;
 103         int entry;
 104
 105         /* process the collected SPU PC for the node */
 106
 107         entry = 0;
 108
 109         trace_addr = cbe_read_pm(cpu, trace_address);
 110         while (!(trace_addr & CBE_PM_TRACE_BUF_EMPTY)) {
 111                 /* there is data in the trace buffer to process */
 112                 spu_pc_extract(cpu, entry);
 113
 114                 entry++;
 115
 116                 if (entry >= TRACE_ARRAY_SIZE)
 117                         /* spu_samples is full */
 118                         break;
 119
 120                 trace_addr = cbe_read_pm(cpu, trace_address);
 121         }
 122
 123         return entry;
 124 }
 125
 126
 127 static enum hrtimer_restart profile_spus(struct hrtimer *timer)
 128 {
 129         ktime_t kt;
 130         int cpu, node, k, num_samples, spu_num;
 131
 132         if (!spu_prof_running)
 133                 goto stop;
 134
 135         for_each_online_cpu(cpu) {
 136                 if (cbe_get_hw_thread_id(cpu))
 137                         continue;
 138
 139                 node = cbe_cpu_to_node(cpu);
 140
 141                 /* There should only be one kernel thread at a time processing
 142                  * the samples.  In the very unlikely case that the processing
 143                  * is taking a very long time and multiple kernel threads are
 144                  * started to process the samples.  Make sure only one kernel
 145                  * thread is working on the samples array at a time.  The
 146                  * sample array must be loaded and then processed for a given
 147                  * cpu.  The sample array is not per cpu.
 148                  */
 149                 spin_lock_irqsave(&sample_array_lock,
 150                                   sample_array_lock_flags);
 151                 num_samples = cell_spu_pc_collection(cpu);
 152
 153                 if (num_samples == 0) {
 154                         spin_unlock_irqrestore(&sample_array_lock,
 155                                                sample_array_lock_flags);
 156                         continue;
 157                 }
 158
 159                 for (k = 0; k < SPUS_PER_NODE; k++) {
 160                         spu_num = k + (node * SPUS_PER_NODE);
 161                         spu_sync_buffer(spu_num,
 162                                         samples + (k * TRACE_ARRAY_SIZE),
 163                                         num_samples);
 164                 }
 165
 166                 spin_unlock_irqrestore(&sample_array_lock,
 167                                        sample_array_lock_flags);
 168
 169         }
 170         smp_wmb();      /* insure spu event buffer updates are written */
 171                         /* don't want events intermingled... */
 172
 173         kt = ktime_set(0, profiling_interval);
 174         if (!spu_prof_running)
 175                 goto stop;
 176         hrtimer_forward(timer, timer->base->get_time(), kt);
 177         return HRTIMER_RESTART;
 178
 179  stop:
 180         printk(KERN_INFO "SPU_PROF: spu-prof timer ending\n");
 181         return HRTIMER_NORESTART;
 182 }
 183
 184 static struct hrtimer timer;
 185 /*
 186  * Entry point for SPU profiling.
 187  * NOTE:  SPU profiling is done system-wide, not per-CPU.
 188  *
 189  * cycles_reset is the count value specified by the user when
 190  * setting up OProfile to count SPU_CYCLES.
 191  */
 192 int start_spu_profiling(unsigned int cycles_reset)
 193 {
 194         ktime_t kt;
 195
 196         pr_debug("timer resolution: %lu\n", TICK_NSEC);
 197         kt = ktime_set(0, profiling_interval);
 198         hrtimer_init(&timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 199         timer.expires = kt;
 200         timer.function = profile_spus;
 201
 202         /* Allocate arrays for collecting SPU PC samples */
 203         samples = kzalloc(SPUS_PER_NODE *
 204                           TRACE_ARRAY_SIZE * sizeof(u32), GFP_KERNEL);
 205
 206         if (!samples)
 207                 return -ENOMEM;
 208
 209         spu_prof_running = 1;
 210         hrtimer_start(&timer, kt, HRTIMER_MODE_REL);
 211
 212         return 0;
 213 }
 214
 215 void stop_spu_profiling(void)
 216 {
 217         spu_prof_running = 0;
 218         hrtimer_cancel(&timer);
 219         kfree(samples);
 220         pr_debug("SPU_PROF: stop_spu_profiling issued\n");
 221 }