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 <asm/time.h>
20 #include "pr_util.h"
21
22 #define SCALE_SHIFT 14
23
24 static u32 *samples;
25
26 /* spu_prof_running is a flag used to indicate if spu profiling is enabled
27 * or not. It is set by the routines start_spu_profiling_cycles() and
28 * start_spu_profiling_events(). The flag is cleared by the routines
29 * stop_spu_profiling_cycles() and stop_spu_profiling_events(). These
30 * routines are called via global_start() and global_stop() which are called in
31 * op_powerpc_start() and op_powerpc_stop(). These routines are called once
32 * per system as a result of the user starting/stopping oprofile. Hence, only
33 * one CPU per user at a time will be changing the value of spu_prof_running.
34 * In general, OProfile does not protect against multiple users trying to run
35 * OProfile at a time.
36 */
37 int spu_prof_running;
38 static unsigned int profiling_interval;
39
40 #define NUM_SPU_BITS_TRBUF 16
41 #define SPUS_PER_TB_ENTRY 4
42
43 #define SPU_PC_MASK 0xFFFF
44
45 DEFINE_SPINLOCK(oprof_spu_smpl_arry_lck);
46 unsigned long oprof_spu_smpl_arry_lck_flags;
47
set_spu_profiling_frequency(unsigned int freq_khz,unsigned int cycles_reset)48 void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_reset)
49 {
50 unsigned long ns_per_cyc;
51
52 if (!freq_khz)
53 freq_khz = ppc_proc_freq/1000;
54
55 /* To calculate a timeout in nanoseconds, the basic
56 * formula is ns = cycles_reset * (NSEC_PER_SEC / cpu frequency).
57 * To avoid floating point math, we use the scale math
58 * technique as described in linux/jiffies.h. We use
59 * a scale factor of SCALE_SHIFT, which provides 4 decimal places
60 * of precision. This is close enough for the purpose at hand.
61 *
62 * The value of the timeout should be small enough that the hw
63 * trace buffer will not get more than about 1/3 full for the
64 * maximum user specified (the LFSR value) hw sampling frequency.
65 * This is to ensure the trace buffer will never fill even if the
66 * kernel thread scheduling varies under a heavy system load.
67 */
68
69 ns_per_cyc = (USEC_PER_SEC << SCALE_SHIFT)/freq_khz;
70 profiling_interval = (ns_per_cyc * cycles_reset) >> SCALE_SHIFT;
71
72 }
73
74 /*
75 * Extract SPU PC from trace buffer entry
76 */
spu_pc_extract(int cpu,int entry)77 static void spu_pc_extract(int cpu, int entry)
78 {
79 /* the trace buffer is 128 bits */
80 u64 trace_buffer[2];
81 u64 spu_mask;
82 int spu;
83
84 spu_mask = SPU_PC_MASK;
85
86 /* Each SPU PC is 16 bits; hence, four spus in each of
87 * the two 64-bit buffer entries that make up the
88 * 128-bit trace_buffer entry. Process two 64-bit values
89 * simultaneously.
90 * trace[0] SPU PC contents are: 0 1 2 3
91 * trace[1] SPU PC contents are: 4 5 6 7
92 */
93
94 cbe_read_trace_buffer(cpu, trace_buffer);
95
96 for (spu = SPUS_PER_TB_ENTRY-1; spu >= 0; spu--) {
97 /* spu PC trace entry is upper 16 bits of the
98 * 18 bit SPU program counter
99 */
100 samples[spu * TRACE_ARRAY_SIZE + entry]
101 = (spu_mask & trace_buffer[0]) << 2;
102 samples[(spu + SPUS_PER_TB_ENTRY) * TRACE_ARRAY_SIZE + entry]
103 = (spu_mask & trace_buffer[1]) << 2;
104
105 trace_buffer[0] = trace_buffer[0] >> NUM_SPU_BITS_TRBUF;
106 trace_buffer[1] = trace_buffer[1] >> NUM_SPU_BITS_TRBUF;
107 }
108 }
109
cell_spu_pc_collection(int cpu)110 static int cell_spu_pc_collection(int cpu)
111 {
112 u32 trace_addr;
113 int entry;
114
115 /* process the collected SPU PC for the node */
116
117 entry = 0;
118
119 trace_addr = cbe_read_pm(cpu, trace_address);
120 while (!(trace_addr & CBE_PM_TRACE_BUF_EMPTY)) {
121 /* there is data in the trace buffer to process */
122 spu_pc_extract(cpu, entry);
123
124 entry++;
125
126 if (entry >= TRACE_ARRAY_SIZE)
127 /* spu_samples is full */
128 break;
129
130 trace_addr = cbe_read_pm(cpu, trace_address);
131 }
132
133 return entry;
134 }
135
136
profile_spus(struct hrtimer * timer)137 static enum hrtimer_restart profile_spus(struct hrtimer *timer)
138 {
139 ktime_t kt;
140 int cpu, node, k, num_samples, spu_num;
141
142 if (!spu_prof_running)
143 goto stop;
144
145 for_each_online_cpu(cpu) {
146 if (cbe_get_hw_thread_id(cpu))
147 continue;
148
149 node = cbe_cpu_to_node(cpu);
150
151 /* There should only be one kernel thread at a time processing
152 * the samples. In the very unlikely case that the processing
153 * is taking a very long time and multiple kernel threads are
154 * started to process the samples. Make sure only one kernel
155 * thread is working on the samples array at a time. The
156 * sample array must be loaded and then processed for a given
157 * cpu. The sample array is not per cpu.
158 */
159 spin_lock_irqsave(&oprof_spu_smpl_arry_lck,
160 oprof_spu_smpl_arry_lck_flags);
161 num_samples = cell_spu_pc_collection(cpu);
162
163 if (num_samples == 0) {
164 spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck,
165 oprof_spu_smpl_arry_lck_flags);
166 continue;
167 }
168
169 for (k = 0; k < SPUS_PER_NODE; k++) {
170 spu_num = k + (node * SPUS_PER_NODE);
171 spu_sync_buffer(spu_num,
172 samples + (k * TRACE_ARRAY_SIZE),
173 num_samples);
174 }
175
176 spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck,
177 oprof_spu_smpl_arry_lck_flags);
178
179 }
180 smp_wmb(); /* insure spu event buffer updates are written */
181 /* don't want events intermingled... */
182
183 kt = ktime_set(0, profiling_interval);
184 if (!spu_prof_running)
185 goto stop;
186 hrtimer_forward(timer, timer->base->get_time(), kt);
187 return HRTIMER_RESTART;
188
189 stop:
190 printk(KERN_INFO "SPU_PROF: spu-prof timer ending\n");
191 return HRTIMER_NORESTART;
192 }
193
194 static struct hrtimer timer;
195 /*
196 * Entry point for SPU cycle profiling.
197 * NOTE: SPU profiling is done system-wide, not per-CPU.
198 *
199 * cycles_reset is the count value specified by the user when
200 * setting up OProfile to count SPU_CYCLES.
201 */
start_spu_profiling_cycles(unsigned int cycles_reset)202 int start_spu_profiling_cycles(unsigned int cycles_reset)
203 {
204 ktime_t kt;
205
206 pr_debug("timer resolution: %lu\n", TICK_NSEC);
207 kt = ktime_set(0, profiling_interval);
208 hrtimer_init(&timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
209 hrtimer_set_expires(&timer, kt);
210 timer.function = profile_spus;
211
212 /* Allocate arrays for collecting SPU PC samples */
213 samples = kzalloc(SPUS_PER_NODE *
214 TRACE_ARRAY_SIZE * sizeof(u32), GFP_KERNEL);
215
216 if (!samples)
217 return -ENOMEM;
218
219 spu_prof_running = 1;
220 hrtimer_start(&timer, kt, HRTIMER_MODE_REL);
221 schedule_delayed_work(&spu_work, DEFAULT_TIMER_EXPIRE);
222
223 return 0;
224 }
225
226 /*
227 * Entry point for SPU event profiling.
228 * NOTE: SPU profiling is done system-wide, not per-CPU.
229 *
230 * cycles_reset is the count value specified by the user when
231 * setting up OProfile to count SPU_CYCLES.
232 */
start_spu_profiling_events(void)233 void start_spu_profiling_events(void)
234 {
235 spu_prof_running = 1;
236 schedule_delayed_work(&spu_work, DEFAULT_TIMER_EXPIRE);
237
238 return;
239 }
240
stop_spu_profiling_cycles(void)241 void stop_spu_profiling_cycles(void)
242 {
243 spu_prof_running = 0;
244 hrtimer_cancel(&timer);
245 kfree(samples);
246 pr_debug("SPU_PROF: stop_spu_profiling_cycles issued\n");
247 }
248
stop_spu_profiling_events(void)249 void stop_spu_profiling_events(void)
250 {
251 spu_prof_running = 0;
252 }
253