2 * Cell Broadband Engine OProfile Support
4 * (C) Copyright IBM Corporation 2006
6 * Authors: Maynard Johnson <maynardj@us.ibm.com>
7 * Carl Love <carll@us.ibm.com>
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.
15 #include <linux/hrtimer.h>
16 #include <linux/smp.h>
17 #include <linux/slab.h>
18 #include <asm/cell-pmu.h>
21 #define TRACE_ARRAY_SIZE 1024
22 #define SCALE_SHIFT 14
26 static int spu_prof_running;
27 static unsigned int profiling_interval;
29 #define NUM_SPU_BITS_TRBUF 16
30 #define SPUS_PER_TB_ENTRY 4
31 #define SPUS_PER_NODE 8
33 #define SPU_PC_MASK 0xFFFF
35 static DEFINE_SPINLOCK(sample_array_lock);
36 unsigned long sample_array_lock_flags;
38 void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_reset)
40 unsigned long ns_per_cyc;
43 freq_khz = ppc_proc_freq/1000;
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.
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.
59 ns_per_cyc = (USEC_PER_SEC << SCALE_SHIFT)/freq_khz;
60 profiling_interval = (ns_per_cyc * cycles_reset) >> SCALE_SHIFT;
65 * Extract SPU PC from trace buffer entry
67 static void spu_pc_extract(int cpu, int entry)
69 /* the trace buffer is 128 bits */
74 spu_mask = SPU_PC_MASK;
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
80 * trace[0] SPU PC contents are: 0 1 2 3
81 * trace[1] SPU PC contents are: 4 5 6 7
84 cbe_read_trace_buffer(cpu, trace_buffer);
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
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;
95 trace_buffer[0] = trace_buffer[0] >> NUM_SPU_BITS_TRBUF;
96 trace_buffer[1] = trace_buffer[1] >> NUM_SPU_BITS_TRBUF;
100 static int cell_spu_pc_collection(int cpu)
105 /* process the collected SPU PC for the node */
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);
116 if (entry >= TRACE_ARRAY_SIZE)
117 /* spu_samples is full */
120 trace_addr = cbe_read_pm(cpu, trace_address);
127 static enum hrtimer_restart profile_spus(struct hrtimer *timer)
130 int cpu, node, k, num_samples, spu_num;
132 if (!spu_prof_running)
135 for_each_online_cpu(cpu) {
136 if (cbe_get_hw_thread_id(cpu))
139 node = cbe_cpu_to_node(cpu);
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.
149 spin_lock_irqsave(&sample_array_lock,
150 sample_array_lock_flags);
151 num_samples = cell_spu_pc_collection(cpu);
153 if (num_samples == 0) {
154 spin_unlock_irqrestore(&sample_array_lock,
155 sample_array_lock_flags);
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),
166 spin_unlock_irqrestore(&sample_array_lock,
167 sample_array_lock_flags);
170 smp_wmb(); /* insure spu event buffer updates are written */
171 /* don't want events intermingled... */
173 kt = ktime_set(0, profiling_interval);
174 if (!spu_prof_running)
176 hrtimer_forward(timer, timer->base->get_time(), kt);
177 return HRTIMER_RESTART;
180 printk(KERN_INFO "SPU_PROF: spu-prof timer ending\n");
181 return HRTIMER_NORESTART;
184 static struct hrtimer timer;
186 * Entry point for SPU profiling.
187 * NOTE: SPU profiling is done system-wide, not per-CPU.
189 * cycles_reset is the count value specified by the user when
190 * setting up OProfile to count SPU_CYCLES.
192 int start_spu_profiling(unsigned int cycles_reset)
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);
200 timer.function = profile_spus;
202 /* Allocate arrays for collecting SPU PC samples */
203 samples = kzalloc(SPUS_PER_NODE *
204 TRACE_ARRAY_SIZE * sizeof(u32), GFP_KERNEL);
209 spu_prof_running = 1;
210 hrtimer_start(&timer, kt, HRTIMER_MODE_REL);
215 void stop_spu_profiling(void)
217 spu_prof_running = 0;
218 hrtimer_cancel(&timer);
220 pr_debug("SPU_PROF: stop_spu_profiling issued\n");