1/**
2 * @file cpu_buffer.c
3 *
4 * @remark Copyright 2002 OProfile authors
5 * @remark Read the file COPYING
6 *
7 * @author John Levon <levon@movementarian.org>
8 *
9 * Each CPU has a local buffer that stores PC value/event
10 * pairs. We also log context switches when we notice them.
11 * Eventually each CPU's buffer is processed into the global
12 * event buffer by sync_buffer().
13 *
14 * We use a local buffer for two reasons: an NMI or similar
15 * interrupt cannot synchronise, and high sampling rates
16 * would lead to catastrophic global synchronisation if
17 * a global buffer was used.
18 */
19
20#include <linux/sched.h>
21#include <linux/oprofile.h>
22#include <linux/vmalloc.h>
23#include <linux/errno.h>
24
25#include "event_buffer.h"
26#include "cpu_buffer.h"
27#include "buffer_sync.h"
28#include "oprof.h"
29
30struct oprofile_cpu_buffer cpu_buffer[NR_CPUS] __cacheline_aligned;
31
32static void wq_sync_buffer(struct work_struct *work);
33
34#define DEFAULT_TIMER_EXPIRE (HZ / 10)
35static int work_enabled;
36
37void free_cpu_buffers(void)
38{
39	int i;
40
41	for_each_online_cpu(i)
42		vfree(cpu_buffer[i].buffer);
43}
44
45int alloc_cpu_buffers(void)
46{
47	int i;
48
49	unsigned long buffer_size = fs_cpu_buffer_size;
50
51	for_each_online_cpu(i) {
52		struct oprofile_cpu_buffer * b = &cpu_buffer[i];
53
54		b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
55			cpu_to_node(i));
56		if (!b->buffer)
57			goto fail;
58
59		b->last_task = NULL;
60		b->last_is_kernel = -1;
61		b->tracing = 0;
62		b->buffer_size = buffer_size;
63		b->tail_pos = 0;
64		b->head_pos = 0;
65		b->sample_received = 0;
66		b->sample_lost_overflow = 0;
67		b->cpu = i;
68		INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
69	}
70	return 0;
71
72fail:
73	free_cpu_buffers();
74	return -ENOMEM;
75}
76
77void start_cpu_work(void)
78{
79	int i;
80
81	work_enabled = 1;
82
83	for_each_online_cpu(i) {
84		struct oprofile_cpu_buffer * b = &cpu_buffer[i];
85
86		/*
87		 * Spread the work by 1 jiffy per cpu so they dont all
88		 * fire at once.
89		 */
90		schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
91	}
92}
93
94void end_cpu_work(void)
95{
96	int i;
97
98	work_enabled = 0;
99
100	for_each_online_cpu(i) {
101		struct oprofile_cpu_buffer * b = &cpu_buffer[i];
102
103		cancel_delayed_work(&b->work);
104	}
105
106	flush_scheduled_work();
107}
108
109/* Resets the cpu buffer to a sane state. */
110void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
111{
112	/* reset these to invalid values; the next sample
113	 * collected will populate the buffer with proper
114	 * values to initialize the buffer
115	 */
116	cpu_buf->last_is_kernel = -1;
117	cpu_buf->last_task = NULL;
118}
119
120/* compute number of available slots in cpu_buffer queue */
121static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
122{
123	unsigned long head = b->head_pos;
124	unsigned long tail = b->tail_pos;
125
126	if (tail > head)
127		return (tail - head) - 1;
128
129	return tail + (b->buffer_size - head) - 1;
130}
131
132static void increment_head(struct oprofile_cpu_buffer * b)
133{
134	unsigned long new_head = b->head_pos + 1;
135
136	/* Ensure anything written to the slot before we
137	 * increment is visible */
138	wmb();
139
140	if (new_head < b->buffer_size)
141		b->head_pos = new_head;
142	else
143		b->head_pos = 0;
144}
145
146static inline void
147add_sample(struct oprofile_cpu_buffer * cpu_buf,
148           unsigned long pc, unsigned long event)
149{
150	struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
151	entry->eip = pc;
152	entry->event = event;
153	increment_head(cpu_buf);
154}
155
156static inline void
157add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
158{
159	add_sample(buffer, ESCAPE_CODE, value);
160}
161
162/* This must be safe from any context. It's safe writing here
163 * because of the head/tail separation of the writer and reader
164 * of the CPU buffer.
165 *
166 * is_kernel is needed because on some architectures you cannot
167 * tell if you are in kernel or user space simply by looking at
168 * pc. We tag this in the buffer by generating kernel enter/exit
169 * events whenever is_kernel changes
170 */
171static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
172		      int is_kernel, unsigned long event)
173{
174	struct task_struct * task;
175
176	cpu_buf->sample_received++;
177
178	if (nr_available_slots(cpu_buf) < 3) {
179		cpu_buf->sample_lost_overflow++;
180		return 0;
181	}
182
183	is_kernel = !!is_kernel;
184
185	task = current;
186
187	/* notice a switch from user->kernel or vice versa */
188	if (cpu_buf->last_is_kernel != is_kernel) {
189		cpu_buf->last_is_kernel = is_kernel;
190		add_code(cpu_buf, is_kernel);
191	}
192
193	/* notice a task switch */
194	if (cpu_buf->last_task != task) {
195		cpu_buf->last_task = task;
196		add_code(cpu_buf, (unsigned long)task);
197	}
198
199	add_sample(cpu_buf, pc, event);
200	return 1;
201}
202
203static int oprofile_begin_trace(struct oprofile_cpu_buffer * cpu_buf)
204{
205	if (nr_available_slots(cpu_buf) < 4) {
206		cpu_buf->sample_lost_overflow++;
207		return 0;
208	}
209
210	add_code(cpu_buf, CPU_TRACE_BEGIN);
211	cpu_buf->tracing = 1;
212	return 1;
213}
214
215static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
216{
217	cpu_buf->tracing = 0;
218}
219
220void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
221				unsigned long event, int is_kernel)
222{
223	struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
224
225	if (!backtrace_depth) {
226		log_sample(cpu_buf, pc, is_kernel, event);
227		return;
228	}
229
230	if (!oprofile_begin_trace(cpu_buf))
231		return;
232
233	/* if log_sample() fail we can't backtrace since we lost the source
234	 * of this event */
235	if (log_sample(cpu_buf, pc, is_kernel, event))
236		oprofile_ops.backtrace(regs, backtrace_depth);
237	oprofile_end_trace(cpu_buf);
238}
239
240void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
241{
242	int is_kernel = !user_mode(regs);
243	unsigned long pc = profile_pc(regs);
244
245	oprofile_add_ext_sample(pc, regs, event, is_kernel);
246}
247
248void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
249{
250	struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
251	log_sample(cpu_buf, pc, is_kernel, event);
252}
253
254void oprofile_add_trace(unsigned long pc)
255{
256	struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
257
258	if (!cpu_buf->tracing)
259		return;
260
261	if (nr_available_slots(cpu_buf) < 1) {
262		cpu_buf->tracing = 0;
263		cpu_buf->sample_lost_overflow++;
264		return;
265	}
266
267	/* broken frame can give an eip with the same value as an escape code,
268	 * abort the trace if we get it */
269	if (pc == ESCAPE_CODE) {
270		cpu_buf->tracing = 0;
271		cpu_buf->backtrace_aborted++;
272		return;
273	}
274
275	add_sample(cpu_buf, pc, 0);
276}
277
278/*
279 * This serves to avoid cpu buffer overflow, and makes sure
280 * the task mortuary progresses
281 *
282 * By using schedule_delayed_work_on and then schedule_delayed_work
283 * we guarantee this will stay on the correct cpu
284 */
285static void wq_sync_buffer(struct work_struct *work)
286{
287	struct oprofile_cpu_buffer * b =
288		container_of(work, struct oprofile_cpu_buffer, work.work);
289	if (b->cpu != smp_processor_id()) {
290		printk("WQ on CPU%d, prefer CPU%d\n",
291		       smp_processor_id(), b->cpu);
292	}
293	sync_buffer(b->cpu);
294
295	/* don't re-add the work if we're shutting down */
296	if (work_enabled)
297		schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
298}
299