trace_hw_branches.c 6.3 KB
Newer Older
1
/*
2
 * h/w branch tracer for x86 based on bts
3
 *
4 5
 * Copyright (C) 2008-2009 Intel Corporation.
 * Markus Metzger <markus.t.metzger@gmail.com>, 2008-2009
6 7 8 9 10 11 12 13
 *
 */

#include <linux/module.h>
#include <linux/fs.h>
#include <linux/debugfs.h>
#include <linux/ftrace.h>
#include <linux/kallsyms.h>
14 15 16
#include <linux/mutex.h>
#include <linux/cpu.h>
#include <linux/smp.h>
17 18 19 20

#include <asm/ds.h>

#include "trace.h"
21
#include "trace_output.h"
22 23 24 25


#define SIZEOF_BTS (1 << 13)

26 27 28 29 30 31 32 33 34 35
/* The tracer mutex protects the below per-cpu tracer array.
   It needs to be held to:
   - start tracing on all cpus
   - stop tracing on all cpus
   - start tracing on a single hotplug cpu
   - stop tracing on a single hotplug cpu
   - read the trace from all cpus
   - read the trace from a single cpu
*/
static DEFINE_MUTEX(bts_tracer_mutex);
36 37 38 39 40 41
static DEFINE_PER_CPU(struct bts_tracer *, tracer);
static DEFINE_PER_CPU(unsigned char[SIZEOF_BTS], buffer);

#define this_tracer per_cpu(tracer, smp_processor_id())
#define this_buffer per_cpu(buffer, smp_processor_id())

42
static int __read_mostly trace_hw_branches_enabled;
43

44 45 46 47 48 49 50

/*
 * Start tracing on the current cpu.
 * The argument is ignored.
 *
 * pre: bts_tracer_mutex must be locked.
 */
51 52
static void bts_trace_start_cpu(void *arg)
{
53 54 55
	if (this_tracer)
		ds_release_bts(this_tracer);

56 57
	this_tracer =
		ds_request_bts(/* task = */ NULL, this_buffer, SIZEOF_BTS,
58 59
			       /* ovfl = */ NULL, /* th = */ (size_t)-1,
			       BTS_KERNEL);
60 61 62 63 64 65 66 67
	if (IS_ERR(this_tracer)) {
		this_tracer = NULL;
		return;
	}
}

static void bts_trace_start(struct trace_array *tr)
{
68
	mutex_lock(&bts_tracer_mutex);
69

70 71
	on_each_cpu(bts_trace_start_cpu, NULL, 1);
	trace_hw_branches_enabled = 1;
72

73
	mutex_unlock(&bts_tracer_mutex);
74 75
}

76 77 78 79 80 81
/*
 * Start tracing on the current cpu.
 * The argument is ignored.
 *
 * pre: bts_tracer_mutex must be locked.
 */
82 83 84 85 86 87 88 89 90 91
static void bts_trace_stop_cpu(void *arg)
{
	if (this_tracer) {
		ds_release_bts(this_tracer);
		this_tracer = NULL;
	}
}

static void bts_trace_stop(struct trace_array *tr)
{
92 93 94 95
	mutex_lock(&bts_tracer_mutex);

	trace_hw_branches_enabled = 0;
	on_each_cpu(bts_trace_stop_cpu, NULL, 1);
96

97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
	mutex_unlock(&bts_tracer_mutex);
}

static int __cpuinit bts_hotcpu_handler(struct notifier_block *nfb,
				     unsigned long action, void *hcpu)
{
	unsigned int cpu = (unsigned long)hcpu;

	mutex_lock(&bts_tracer_mutex);

	if (!trace_hw_branches_enabled)
		goto out;

	switch (action) {
	case CPU_ONLINE:
	case CPU_DOWN_FAILED:
		smp_call_function_single(cpu, bts_trace_start_cpu, NULL, 1);
		break;
	case CPU_DOWN_PREPARE:
116
		smp_call_function_single(cpu, bts_trace_stop_cpu, NULL, 1);
117 118 119 120 121 122
		break;
	}

 out:
	mutex_unlock(&bts_tracer_mutex);
	return NOTIFY_DONE;
123 124
}

125 126 127 128
static struct notifier_block bts_hotcpu_notifier __cpuinitdata = {
	.notifier_call = bts_hotcpu_handler
};

129 130
static int bts_trace_init(struct trace_array *tr)
{
131
	register_hotcpu_notifier(&bts_hotcpu_notifier);
132
	tracing_reset_online_cpus(tr);
133 134 135 136 137
	bts_trace_start(tr);

	return 0;
}

138 139 140 141 142 143
static void bts_trace_reset(struct trace_array *tr)
{
	bts_trace_stop(tr);
	unregister_hotcpu_notifier(&bts_hotcpu_notifier);
}

144 145 146 147 148 149 150 151 152 153 154 155
static void bts_trace_print_header(struct seq_file *m)
{
	seq_puts(m,
		 "# CPU#        FROM                   TO         FUNCTION\n");
	seq_puts(m,
		 "#  |           |                     |             |\n");
}

static enum print_line_t bts_trace_print_line(struct trace_iterator *iter)
{
	struct trace_entry *entry = iter->ent;
	struct trace_seq *seq = &iter->seq;
156
	struct hw_branch_entry *it;
157 158 159

	trace_assign_type(it, entry);

160 161 162 163 164 165 166 167 168
	if (entry->type == TRACE_HW_BRANCHES) {
		if (trace_seq_printf(seq, "%4d  ", entry->cpu) &&
		    trace_seq_printf(seq, "0x%016llx -> 0x%016llx ",
				     it->from, it->to) &&
		    (!it->from ||
		     seq_print_ip_sym(seq, it->from, /* sym_flags = */ 0)) &&
		    trace_seq_printf(seq, "\n"))
			return TRACE_TYPE_HANDLED;
		return TRACE_TYPE_PARTIAL_LINE;;
169 170 171 172
	}
	return TRACE_TYPE_UNHANDLED;
}

173
void trace_hw_branch(struct trace_array *tr, u64 from, u64 to)
174 175
{
	struct ring_buffer_event *event;
176
	struct hw_branch_entry *entry;
177 178
	unsigned long irq1, irq2;
	int cpu;
179

180 181 182 183
	if (unlikely(!tr))
		return;

	if (unlikely(!trace_hw_branches_enabled))
184
		return;
185 186 187 188 189 190 191 192 193

	local_irq_save(irq1);
	cpu = raw_smp_processor_id();
	if (atomic_inc_return(&tr->data[cpu]->disabled) != 1)
		goto out;

	event = ring_buffer_lock_reserve(tr->buffer, sizeof(*entry), &irq2);
	if (!event)
		goto out;
194 195
	entry	= ring_buffer_event_data(event);
	tracing_generic_entry_update(&entry->ent, 0, from);
196
	entry->ent.type = TRACE_HW_BRANCHES;
197
	entry->ent.cpu = cpu;
198 199
	entry->from = from;
	entry->to   = to;
200 201 202 203 204
	ring_buffer_unlock_commit(tr->buffer, event, irq2);

 out:
	atomic_dec(&tr->data[cpu]->disabled);
	local_irq_restore(irq1);
205 206
}

207 208
static void trace_bts_at(struct trace_array *tr,
			 const struct bts_trace *trace, void *at)
209
{
210 211
	struct bts_struct bts;
	int err = 0;
212

213 214
	WARN_ON_ONCE(!trace->read);
	if (!trace->read)
215 216
		return;

217 218 219
	err = trace->read(this_tracer, at, &bts);
	if (err < 0)
		return;
220

221 222 223 224 225
	switch (bts.qualifier) {
	case BTS_BRANCH:
		trace_hw_branch(tr, bts.variant.lbr.from, bts.variant.lbr.to);
		break;
	}
226 227
}

228 229 230 231 232
/*
 * Collect the trace on the current cpu and write it into the ftrace buffer.
 *
 * pre: bts_tracer_mutex must be locked
 */
233 234 235
static void trace_bts_cpu(void *arg)
{
	struct trace_array *tr = (struct trace_array *) arg;
236 237
	const struct bts_trace *trace;
	unsigned char *at;
238 239 240 241

	if (!this_tracer)
		return;

242 243 244
	if (unlikely(atomic_read(&tr->data[raw_smp_processor_id()]->disabled)))
		return;

245 246 247
	ds_suspend_bts(this_tracer);
	trace = ds_read_bts(this_tracer);
	if (!trace)
248 249
		goto out;

250 251 252
	for (at = trace->ds.top; (void *)at < trace->ds.end;
	     at += trace->ds.size)
		trace_bts_at(tr, trace, at);
253

254 255 256
	for (at = trace->ds.begin; (void *)at < trace->ds.top;
	     at += trace->ds.size)
		trace_bts_at(tr, trace, at);
257 258

out:
259
	ds_resume_bts(this_tracer);
260 261 262 263
}

static void trace_bts_prepare(struct trace_iterator *iter)
{
264 265 266
	mutex_lock(&bts_tracer_mutex);

	on_each_cpu(trace_bts_cpu, iter->tr, 1);
267

268
	mutex_unlock(&bts_tracer_mutex);
269 270 271 272
}

struct tracer bts_tracer __read_mostly =
{
273
	.name		= "hw-branch-tracer",
274
	.init		= bts_trace_init,
275
	.reset		= bts_trace_reset,
276 277 278 279 280 281 282 283 284 285 286 287
	.print_header	= bts_trace_print_header,
	.print_line	= bts_trace_print_line,
	.start		= bts_trace_start,
	.stop		= bts_trace_stop,
	.open		= trace_bts_prepare
};

__init static int init_bts_trace(void)
{
	return register_tracer(&bts_tracer);
}
device_initcall(init_bts_trace);