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
27 static unsigned int profiling_interval;
29 #define NUM_SPU_BITS_TRBUF 16
30 #define SPUS_PER_TB_ENTRY 4
32 #define SPU_PC_MASK 0xFFFF
34 static DEFINE_SPINLOCK(sample_array_lock);
35 unsigned long sample_array_lock_flags;
37 void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_reset)
39 unsigned long ns_per_cyc;
42 freq_khz = ppc_proc_freq/1000;
44 /* To calculate a timeout in nanoseconds, the basic
45 * formula is ns = cycles_reset * (NSEC_PER_SEC / cpu frequency).
46 * To avoid floating point math, we use the scale math
47 * technique as described in linux/jiffies.h. We use
48 * a scale factor of SCALE_SHIFT, which provides 4 decimal places
49 * of precision. This is close enough for the purpose at hand.
51 * The value of the timeout should be small enough that the hw
52 * trace buffer will not get more than about 1/3 full for the
53 * maximum user specified (the LFSR value) hw sampling frequency.
54 * This is to ensure the trace buffer will never fill even if the
55 * kernel thread scheduling varies under a heavy system load.
58 ns_per_cyc = (USEC_PER_SEC << SCALE_SHIFT)/freq_khz;
59 profiling_interval = (ns_per_cyc * cycles_reset) >> SCALE_SHIFT;
64 * Extract SPU PC from trace buffer entry
66 static void spu_pc_extract(int cpu, int entry)
68 /* the trace buffer is 128 bits */
73 spu_mask = SPU_PC_MASK;
75 /* Each SPU PC is 16 bits; hence, four spus in each of
76 * the two 64-bit buffer entries that make up the
77 * 128-bit trace_buffer entry. Process two 64-bit values
79 * trace[0] SPU PC contents are: 0 1 2 3
80 * trace[1] SPU PC contents are: 4 5 6 7
83 cbe_read_trace_buffer(cpu, trace_buffer);
85 for (spu = SPUS_PER_TB_ENTRY-1; spu >= 0; spu--) {
86 /* spu PC trace entry is upper 16 bits of the
87 * 18 bit SPU program counter
89 samples[spu * TRACE_ARRAY_SIZE + entry]
90 = (spu_mask & trace_buffer[0]) << 2;
91 samples[(spu + SPUS_PER_TB_ENTRY) * TRACE_ARRAY_SIZE + entry]
92 = (spu_mask & trace_buffer[1]) << 2;
94 trace_buffer[0] = trace_buffer[0] >> NUM_SPU_BITS_TRBUF;
95 trace_buffer[1] = trace_buffer[1] >> NUM_SPU_BITS_TRBUF;
99 static int cell_spu_pc_collection(int cpu)
104 /* process the collected SPU PC for the node */
108 trace_addr = cbe_read_pm(cpu, trace_address);
109 while (!(trace_addr & CBE_PM_TRACE_BUF_EMPTY)) {
110 /* there is data in the trace buffer to process */
111 spu_pc_extract(cpu, entry);
115 if (entry >= TRACE_ARRAY_SIZE)
116 /* spu_samples is full */
119 trace_addr = cbe_read_pm(cpu, trace_address);
126 static enum hrtimer_restart profile_spus(struct hrtimer *timer)
129 int cpu, node, k, num_samples, spu_num;
131 if (!spu_prof_running)
134 for_each_online_cpu(cpu) {
135 if (cbe_get_hw_thread_id(cpu))
138 node = cbe_cpu_to_node(cpu);
140 /* There should only be one kernel thread at a time processing
141 * the samples. In the very unlikely case that the processing
142 * is taking a very long time and multiple kernel threads are
143 * started to process the samples. Make sure only one kernel
144 * thread is working on the samples array at a time. The
145 * sample array must be loaded and then processed for a given
146 * cpu. The sample array is not per cpu.
148 spin_lock_irqsave(&sample_array_lock,
149 sample_array_lock_flags);
150 num_samples = cell_spu_pc_collection(cpu);
152 if (num_samples == 0) {
153 spin_unlock_irqrestore(&sample_array_lock,
154 sample_array_lock_flags);
158 for (k = 0; k < SPUS_PER_NODE; k++) {
159 spu_num = k + (node * SPUS_PER_NODE);
160 spu_sync_buffer(spu_num,
161 samples + (k * TRACE_ARRAY_SIZE),
165 spin_unlock_irqrestore(&sample_array_lock,
166 sample_array_lock_flags);
169 smp_wmb(); /* insure spu event buffer updates are written */
170 /* don't want events intermingled... */
172 kt = ktime_set(0, profiling_interval);
173 if (!spu_prof_running)
175 hrtimer_forward(timer, timer->base->get_time(), kt);
176 return HRTIMER_RESTART;
179 printk(KERN_INFO "SPU_PROF: spu-prof timer ending\n");
180 return HRTIMER_NORESTART;
183 static struct hrtimer timer;
185 * Entry point for SPU profiling.
186 * NOTE: SPU profiling is done system-wide, not per-CPU.
188 * cycles_reset is the count value specified by the user when
189 * setting up OProfile to count SPU_CYCLES.
191 int start_spu_profiling(unsigned int cycles_reset)
195 pr_debug("timer resolution: %lu\n", TICK_NSEC);
196 kt = ktime_set(0, profiling_interval);
197 hrtimer_init(&timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
198 hrtimer_set_expires(&timer, kt);
199 timer.function = profile_spus;
201 /* Allocate arrays for collecting SPU PC samples */
202 samples = kzalloc(SPUS_PER_NODE *
203 TRACE_ARRAY_SIZE * sizeof(u32), GFP_KERNEL);
208 spu_prof_running = 1;
209 hrtimer_start(&timer, kt, HRTIMER_MODE_REL);
210 schedule_delayed_work(&spu_work, DEFAULT_TIMER_EXPIRE);
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");