1/**
2 * @file cpu_buffer.c
3 *
4 * @remark Copyright 2002-2009 OProfile authors
5 * @remark Read the file COPYING
6 *
7 * @author John Levon <levon@movementarian.org>
8 * @author Barry Kasindorf <barry.kasindorf@amd.com>
9 * @author Robert Richter <robert.richter@amd.com>
10 *
11 * Each CPU has a local buffer that stores PC value/event
12 * pairs. We also log context switches when we notice them.
13 * Eventually each CPU's buffer is processed into the global
14 * event buffer by sync_buffer().
15 *
16 * We use a local buffer for two reasons: an NMI or similar
17 * interrupt cannot synchronise, and high sampling rates
18 * would lead to catastrophic global synchronisation if
19 * a global buffer was used.
20 */
21
22#include <linux/sched.h>
23#include <linux/oprofile.h>
24#include <linux/errno.h>
25
26#include "event_buffer.h"
27#include "cpu_buffer.h"
28#include "buffer_sync.h"
29#include "oprof.h"
30
31#define OP_BUFFER_FLAGS	0
32
33static struct ring_buffer *op_ring_buffer;
34DEFINE_PER_CPU(struct oprofile_cpu_buffer, op_cpu_buffer);
35
36static void wq_sync_buffer(struct work_struct *work);
37
38#define DEFAULT_TIMER_EXPIRE (HZ / 10)
39static int work_enabled;
40
41unsigned long oprofile_get_cpu_buffer_size(void)
42{
43	return oprofile_cpu_buffer_size;
44}
45
46void oprofile_cpu_buffer_inc_smpl_lost(void)
47{
48	struct oprofile_cpu_buffer *cpu_buf = this_cpu_ptr(&op_cpu_buffer);
49
50	cpu_buf->sample_lost_overflow++;
51}
52
53void free_cpu_buffers(void)
54{
55	if (op_ring_buffer)
56		ring_buffer_free(op_ring_buffer);
57	op_ring_buffer = NULL;
58}
59
60#define RB_EVENT_HDR_SIZE 4
61
62int alloc_cpu_buffers(void)
63{
64	int i;
65
66	unsigned long buffer_size = oprofile_cpu_buffer_size;
67	unsigned long byte_size = buffer_size * (sizeof(struct op_sample) +
68						 RB_EVENT_HDR_SIZE);
69
70	op_ring_buffer = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS);
71	if (!op_ring_buffer)
72		goto fail;
73
74	for_each_possible_cpu(i) {
75		struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
76
77		b->last_task = NULL;
78		b->last_is_kernel = -1;
79		b->tracing = 0;
80		b->buffer_size = buffer_size;
81		b->sample_received = 0;
82		b->sample_lost_overflow = 0;
83		b->backtrace_aborted = 0;
84		b->sample_invalid_eip = 0;
85		b->cpu = i;
86		INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
87	}
88	return 0;
89
90fail:
91	free_cpu_buffers();
92	return -ENOMEM;
93}
94
95void start_cpu_work(void)
96{
97	int i;
98
99	work_enabled = 1;
100
101	for_each_online_cpu(i) {
102		struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
103
104		/*
105		 * Spread the work by 1 jiffy per cpu so they dont all
106		 * fire at once.
107		 */
108		schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
109	}
110}
111
112void end_cpu_work(void)
113{
114	work_enabled = 0;
115}
116
117void flush_cpu_work(void)
118{
119	int i;
120
121	for_each_online_cpu(i) {
122		struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
123
124		/* these works are per-cpu, no need for flush_sync */
125		flush_delayed_work(&b->work);
126	}
127}
128
129/*
130 * This function prepares the cpu buffer to write a sample.
131 *
132 * Struct op_entry is used during operations on the ring buffer while
133 * struct op_sample contains the data that is stored in the ring
134 * buffer. Struct entry can be uninitialized. The function reserves a
135 * data array that is specified by size. Use
136 * op_cpu_buffer_write_commit() after preparing the sample. In case of
137 * errors a null pointer is returned, otherwise the pointer to the
138 * sample.
139 *
140 */
141struct op_sample
142*op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size)
143{
144	entry->event = ring_buffer_lock_reserve
145		(op_ring_buffer, sizeof(struct op_sample) +
146		 size * sizeof(entry->sample->data[0]));
147	if (!entry->event)
148		return NULL;
149	entry->sample = ring_buffer_event_data(entry->event);
150	entry->size = size;
151	entry->data = entry->sample->data;
152
153	return entry->sample;
154}
155
156int op_cpu_buffer_write_commit(struct op_entry *entry)
157{
158	return ring_buffer_unlock_commit(op_ring_buffer, entry->event);
159}
160
161struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu)
162{
163	struct ring_buffer_event *e;
164	e = ring_buffer_consume(op_ring_buffer, cpu, NULL, NULL);
165	if (!e)
166		return NULL;
167
168	entry->event = e;
169	entry->sample = ring_buffer_event_data(e);
170	entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample))
171		/ sizeof(entry->sample->data[0]);
172	entry->data = entry->sample->data;
173	return entry->sample;
174}
175
176unsigned long op_cpu_buffer_entries(int cpu)
177{
178	return ring_buffer_entries_cpu(op_ring_buffer, cpu);
179}
180
181static int
182op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace,
183	    int is_kernel, struct task_struct *task)
184{
185	struct op_entry entry;
186	struct op_sample *sample;
187	unsigned long flags;
188	int size;
189
190	flags = 0;
191
192	if (backtrace)
193		flags |= TRACE_BEGIN;
194
195	/* notice a switch from user->kernel or vice versa */
196	is_kernel = !!is_kernel;
197	if (cpu_buf->last_is_kernel != is_kernel) {
198		cpu_buf->last_is_kernel = is_kernel;
199		flags |= KERNEL_CTX_SWITCH;
200		if (is_kernel)
201			flags |= IS_KERNEL;
202	}
203
204	/* notice a task switch */
205	if (cpu_buf->last_task != task) {
206		cpu_buf->last_task = task;
207		flags |= USER_CTX_SWITCH;
208	}
209
210	if (!flags)
211		/* nothing to do */
212		return 0;
213
214	if (flags & USER_CTX_SWITCH)
215		size = 1;
216	else
217		size = 0;
218
219	sample = op_cpu_buffer_write_reserve(&entry, size);
220	if (!sample)
221		return -ENOMEM;
222
223	sample->eip = ESCAPE_CODE;
224	sample->event = flags;
225
226	if (size)
227		op_cpu_buffer_add_data(&entry, (unsigned long)task);
228
229	op_cpu_buffer_write_commit(&entry);
230
231	return 0;
232}
233
234static inline int
235op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
236	      unsigned long pc, unsigned long event)
237{
238	struct op_entry entry;
239	struct op_sample *sample;
240
241	sample = op_cpu_buffer_write_reserve(&entry, 0);
242	if (!sample)
243		return -ENOMEM;
244
245	sample->eip = pc;
246	sample->event = event;
247
248	return op_cpu_buffer_write_commit(&entry);
249}
250
251/*
252 * This must be safe from any context.
253 *
254 * is_kernel is needed because on some architectures you cannot
255 * tell if you are in kernel or user space simply by looking at
256 * pc. We tag this in the buffer by generating kernel enter/exit
257 * events whenever is_kernel changes
258 */
259static int
260log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
261	   unsigned long backtrace, int is_kernel, unsigned long event,
262	   struct task_struct *task)
263{
264	struct task_struct *tsk = task ? task : current;
265	cpu_buf->sample_received++;
266
267	if (pc == ESCAPE_CODE) {
268		cpu_buf->sample_invalid_eip++;
269		return 0;
270	}
271
272	if (op_add_code(cpu_buf, backtrace, is_kernel, tsk))
273		goto fail;
274
275	if (op_add_sample(cpu_buf, pc, event))
276		goto fail;
277
278	return 1;
279
280fail:
281	cpu_buf->sample_lost_overflow++;
282	return 0;
283}
284
285static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
286{
287	cpu_buf->tracing = 1;
288}
289
290static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
291{
292	cpu_buf->tracing = 0;
293}
294
295static inline void
296__oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
297			  unsigned long event, int is_kernel,
298			  struct task_struct *task)
299{
300	struct oprofile_cpu_buffer *cpu_buf = this_cpu_ptr(&op_cpu_buffer);
301	unsigned long backtrace = oprofile_backtrace_depth;
302
303	/*
304	 * if log_sample() fail we can't backtrace since we lost the
305	 * source of this event
306	 */
307	if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event, task))
308		/* failed */
309		return;
310
311	if (!backtrace)
312		return;
313
314	oprofile_begin_trace(cpu_buf);
315	oprofile_ops.backtrace(regs, backtrace);
316	oprofile_end_trace(cpu_buf);
317}
318
319void oprofile_add_ext_hw_sample(unsigned long pc, struct pt_regs * const regs,
320				unsigned long event, int is_kernel,
321				struct task_struct *task)
322{
323	__oprofile_add_ext_sample(pc, regs, event, is_kernel, task);
324}
325
326void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
327			     unsigned long event, int is_kernel)
328{
329	__oprofile_add_ext_sample(pc, regs, event, is_kernel, NULL);
330}
331
332void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
333{
334	int is_kernel;
335	unsigned long pc;
336
337	if (likely(regs)) {
338		is_kernel = !user_mode(regs);
339		pc = profile_pc(regs);
340	} else {
341		is_kernel = 0;    /* This value will not be used */
342		pc = ESCAPE_CODE; /* as this causes an early return. */
343	}
344
345	__oprofile_add_ext_sample(pc, regs, event, is_kernel, NULL);
346}
347
348/*
349 * Add samples with data to the ring buffer.
350 *
351 * Use oprofile_add_data(&entry, val) to add data and
352 * oprofile_write_commit(&entry) to commit the sample.
353 */
354void
355oprofile_write_reserve(struct op_entry *entry, struct pt_regs * const regs,
356		       unsigned long pc, int code, int size)
357{
358	struct op_sample *sample;
359	int is_kernel = !user_mode(regs);
360	struct oprofile_cpu_buffer *cpu_buf = this_cpu_ptr(&op_cpu_buffer);
361
362	cpu_buf->sample_received++;
363
364	/* no backtraces for samples with data */
365	if (op_add_code(cpu_buf, 0, is_kernel, current))
366		goto fail;
367
368	sample = op_cpu_buffer_write_reserve(entry, size + 2);
369	if (!sample)
370		goto fail;
371	sample->eip = ESCAPE_CODE;
372	sample->event = 0;		/* no flags */
373
374	op_cpu_buffer_add_data(entry, code);
375	op_cpu_buffer_add_data(entry, pc);
376
377	return;
378
379fail:
380	entry->event = NULL;
381	cpu_buf->sample_lost_overflow++;
382}
383
384int oprofile_add_data(struct op_entry *entry, unsigned long val)
385{
386	if (!entry->event)
387		return 0;
388	return op_cpu_buffer_add_data(entry, val);
389}
390
391int oprofile_add_data64(struct op_entry *entry, u64 val)
392{
393	if (!entry->event)
394		return 0;
395	if (op_cpu_buffer_get_size(entry) < 2)
396		/*
397		 * the function returns 0 to indicate a too small
398		 * buffer, even if there is some space left
399		 */
400		return 0;
401	if (!op_cpu_buffer_add_data(entry, (u32)val))
402		return 0;
403	return op_cpu_buffer_add_data(entry, (u32)(val >> 32));
404}
405
406int oprofile_write_commit(struct op_entry *entry)
407{
408	if (!entry->event)
409		return -EINVAL;
410	return op_cpu_buffer_write_commit(entry);
411}
412
413void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
414{
415	struct oprofile_cpu_buffer *cpu_buf = this_cpu_ptr(&op_cpu_buffer);
416	log_sample(cpu_buf, pc, 0, is_kernel, event, NULL);
417}
418
419void oprofile_add_trace(unsigned long pc)
420{
421	struct oprofile_cpu_buffer *cpu_buf = this_cpu_ptr(&op_cpu_buffer);
422
423	if (!cpu_buf->tracing)
424		return;
425
426	/*
427	 * broken frame can give an eip with the same value as an
428	 * escape code, abort the trace if we get it
429	 */
430	if (pc == ESCAPE_CODE)
431		goto fail;
432
433	if (op_add_sample(cpu_buf, pc, 0))
434		goto fail;
435
436	return;
437fail:
438	cpu_buf->tracing = 0;
439	cpu_buf->backtrace_aborted++;
440	return;
441}
442
443/*
444 * This serves to avoid cpu buffer overflow, and makes sure
445 * the task mortuary progresses
446 *
447 * By using schedule_delayed_work_on and then schedule_delayed_work
448 * we guarantee this will stay on the correct cpu
449 */
450static void wq_sync_buffer(struct work_struct *work)
451{
452	struct oprofile_cpu_buffer *b =
453		container_of(work, struct oprofile_cpu_buffer, work.work);
454	if (b->cpu != smp_processor_id() && !cpu_online(b->cpu)) {
455		cancel_delayed_work(&b->work);
456		return;
457	}
458	sync_buffer(b->cpu);
459
460	/* don't re-add the work if we're shutting down */
461	if (work_enabled)
462		schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
463}
464