oprofile: remove nr_available_slots()
[deliverable/linux.git] / drivers / oprofile / cpu_buffer.c
CommitLineData
1da177e4
LT
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>
345c2573 8 * @author Barry Kasindorf <barry.kasindorf@amd.com>
1da177e4
LT
9 *
10 * Each CPU has a local buffer that stores PC value/event
11 * pairs. We also log context switches when we notice them.
12 * Eventually each CPU's buffer is processed into the global
13 * event buffer by sync_buffer().
14 *
15 * We use a local buffer for two reasons: an NMI or similar
16 * interrupt cannot synchronise, and high sampling rates
17 * would lead to catastrophic global synchronisation if
18 * a global buffer was used.
19 */
20
21#include <linux/sched.h>
22#include <linux/oprofile.h>
23#include <linux/vmalloc.h>
24#include <linux/errno.h>
6a18037d 25
1da177e4
LT
26#include "event_buffer.h"
27#include "cpu_buffer.h"
28#include "buffer_sync.h"
29#include "oprof.h"
30
6dad828b
RR
31#define OP_BUFFER_FLAGS 0
32
33/*
34 * Read and write access is using spin locking. Thus, writing to the
35 * buffer by NMI handler (x86) could occur also during critical
36 * sections when reading the buffer. To avoid this, there are 2
37 * buffers for independent read and write access. Read access is in
38 * process context only, write access only in the NMI handler. If the
39 * read buffer runs empty, both buffers are swapped atomically. There
40 * is potentially a small window during swapping where the buffers are
41 * disabled and samples could be lost.
42 *
43 * Using 2 buffers is a little bit overhead, but the solution is clear
44 * and does not require changes in the ring buffer implementation. It
45 * can be changed to a single buffer solution when the ring buffer
46 * access is implemented as non-locking atomic code.
47 */
48struct ring_buffer *op_ring_buffer_read;
49struct ring_buffer *op_ring_buffer_write;
8b8b4988 50DEFINE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
1da177e4 51
c4028958 52static void wq_sync_buffer(struct work_struct *work);
1da177e4
LT
53
54#define DEFAULT_TIMER_EXPIRE (HZ / 10)
55static int work_enabled;
56
57void free_cpu_buffers(void)
58{
6dad828b
RR
59 if (op_ring_buffer_read)
60 ring_buffer_free(op_ring_buffer_read);
61 op_ring_buffer_read = NULL;
62 if (op_ring_buffer_write)
63 ring_buffer_free(op_ring_buffer_write);
64 op_ring_buffer_write = NULL;
1da177e4 65}
77933d72 66
a5598ca0
CL
67unsigned long oprofile_get_cpu_buffer_size(void)
68{
69 return fs_cpu_buffer_size;
70}
71
72void oprofile_cpu_buffer_inc_smpl_lost(void)
73{
74 struct oprofile_cpu_buffer *cpu_buf
75 = &__get_cpu_var(cpu_buffer);
76
77 cpu_buf->sample_lost_overflow++;
78}
79
1da177e4
LT
80int alloc_cpu_buffers(void)
81{
82 int i;
6a18037d 83
1da177e4 84 unsigned long buffer_size = fs_cpu_buffer_size;
6a18037d 85
6dad828b
RR
86 op_ring_buffer_read = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
87 if (!op_ring_buffer_read)
88 goto fail;
89 op_ring_buffer_write = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
90 if (!op_ring_buffer_write)
91 goto fail;
92
4bd9b9dc 93 for_each_possible_cpu(i) {
608dfddd 94 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
6a18037d 95
1da177e4
LT
96 b->last_task = NULL;
97 b->last_is_kernel = -1;
98 b->tracing = 0;
99 b->buffer_size = buffer_size;
100 b->tail_pos = 0;
101 b->head_pos = 0;
102 b->sample_received = 0;
103 b->sample_lost_overflow = 0;
df9d177a
PE
104 b->backtrace_aborted = 0;
105 b->sample_invalid_eip = 0;
1da177e4 106 b->cpu = i;
c4028958 107 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
1da177e4
LT
108 }
109 return 0;
110
111fail:
112 free_cpu_buffers();
113 return -ENOMEM;
114}
1da177e4
LT
115
116void start_cpu_work(void)
117{
118 int i;
119
120 work_enabled = 1;
121
122 for_each_online_cpu(i) {
608dfddd 123 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
1da177e4
LT
124
125 /*
126 * Spread the work by 1 jiffy per cpu so they dont all
127 * fire at once.
128 */
129 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
130 }
131}
132
1da177e4
LT
133void end_cpu_work(void)
134{
135 int i;
136
137 work_enabled = 0;
138
139 for_each_online_cpu(i) {
608dfddd 140 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
1da177e4
LT
141
142 cancel_delayed_work(&b->work);
143 }
144
145 flush_scheduled_work();
146}
147
77933d72 148static inline void
25ad2913 149add_sample(struct oprofile_cpu_buffer *cpu_buf,
6a18037d 150 unsigned long pc, unsigned long event)
1da177e4 151{
6dad828b
RR
152 struct op_entry entry;
153
154 if (cpu_buffer_write_entry(&entry))
155 goto Error;
156
157 entry.sample->eip = pc;
158 entry.sample->event = event;
159
160 if (cpu_buffer_write_commit(&entry))
161 goto Error;
162
163 return;
164
165Error:
166 cpu_buf->sample_lost_overflow++;
167 return;
1da177e4
LT
168}
169
77933d72 170static inline void
25ad2913 171add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
1da177e4
LT
172{
173 add_sample(buffer, ESCAPE_CODE, value);
174}
175
1da177e4
LT
176/* This must be safe from any context. It's safe writing here
177 * because of the head/tail separation of the writer and reader
178 * of the CPU buffer.
179 *
180 * is_kernel is needed because on some architectures you cannot
181 * tell if you are in kernel or user space simply by looking at
182 * pc. We tag this in the buffer by generating kernel enter/exit
183 * events whenever is_kernel changes
184 */
25ad2913 185static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
1da177e4
LT
186 int is_kernel, unsigned long event)
187{
25ad2913 188 struct task_struct *task;
1da177e4
LT
189
190 cpu_buf->sample_received++;
191
df9d177a
PE
192 if (pc == ESCAPE_CODE) {
193 cpu_buf->sample_invalid_eip++;
194 return 0;
195 }
196
1da177e4
LT
197 is_kernel = !!is_kernel;
198
199 task = current;
200
201 /* notice a switch from user->kernel or vice versa */
202 if (cpu_buf->last_is_kernel != is_kernel) {
203 cpu_buf->last_is_kernel = is_kernel;
204 add_code(cpu_buf, is_kernel);
205 }
206
207 /* notice a task switch */
208 if (cpu_buf->last_task != task) {
209 cpu_buf->last_task = task;
210 add_code(cpu_buf, (unsigned long)task);
211 }
6a18037d 212
1da177e4
LT
213 add_sample(cpu_buf, pc, event);
214 return 1;
215}
216
345c2573 217static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
1da177e4 218{
1da177e4
LT
219 add_code(cpu_buf, CPU_TRACE_BEGIN);
220 cpu_buf->tracing = 1;
221 return 1;
222}
223
25ad2913 224static void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
1da177e4
LT
225{
226 cpu_buf->tracing = 0;
227}
228
27357716
BR
229void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
230 unsigned long event, int is_kernel)
1da177e4 231{
608dfddd 232 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
1da177e4
LT
233
234 if (!backtrace_depth) {
235 log_sample(cpu_buf, pc, is_kernel, event);
236 return;
237 }
238
239 if (!oprofile_begin_trace(cpu_buf))
240 return;
241
fd13f6c8
RR
242 /*
243 * if log_sample() fail we can't backtrace since we lost the
244 * source of this event
245 */
1da177e4
LT
246 if (log_sample(cpu_buf, pc, is_kernel, event))
247 oprofile_ops.backtrace(regs, backtrace_depth);
248 oprofile_end_trace(cpu_buf);
249}
250
27357716
BR
251void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
252{
253 int is_kernel = !user_mode(regs);
254 unsigned long pc = profile_pc(regs);
255
256 oprofile_add_ext_sample(pc, regs, event, is_kernel);
257}
258
852402cc
RR
259#ifdef CONFIG_OPROFILE_IBS
260
e2fee276
RR
261#define MAX_IBS_SAMPLE_SIZE 14
262
cdc1834d
RR
263void oprofile_add_ibs_sample(struct pt_regs * const regs,
264 unsigned int * const ibs_sample, int ibs_code)
345c2573 265{
e2fee276
RR
266 int is_kernel = !user_mode(regs);
267 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
345c2573
BK
268 struct task_struct *task;
269
270 cpu_buf->sample_received++;
271
345c2573
BK
272 /* notice a switch from user->kernel or vice versa */
273 if (cpu_buf->last_is_kernel != is_kernel) {
274 cpu_buf->last_is_kernel = is_kernel;
275 add_code(cpu_buf, is_kernel);
276 }
277
278 /* notice a task switch */
279 if (!is_kernel) {
280 task = current;
345c2573
BK
281 if (cpu_buf->last_task != task) {
282 cpu_buf->last_task = task;
283 add_code(cpu_buf, (unsigned long)task);
284 }
285 }
286
287 add_code(cpu_buf, ibs_code);
e2fee276
RR
288 add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
289 add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
290 add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
345c2573
BK
291
292 if (ibs_code == IBS_OP_BEGIN) {
e2fee276
RR
293 add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
294 add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
295 add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
345c2573
BK
296 }
297
e2fee276 298 if (backtrace_depth)
345c2573
BK
299 oprofile_ops.backtrace(regs, backtrace_depth);
300}
301
852402cc
RR
302#endif
303
1da177e4
LT
304void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
305{
608dfddd 306 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
1da177e4
LT
307 log_sample(cpu_buf, pc, is_kernel, event);
308}
309
1da177e4
LT
310void oprofile_add_trace(unsigned long pc)
311{
608dfddd 312 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
1da177e4
LT
313
314 if (!cpu_buf->tracing)
315 return;
316
fd13f6c8
RR
317 /*
318 * broken frame can give an eip with the same value as an
319 * escape code, abort the trace if we get it
320 */
1da177e4
LT
321 if (pc == ESCAPE_CODE) {
322 cpu_buf->tracing = 0;
323 cpu_buf->backtrace_aborted++;
324 return;
325 }
326
327 add_sample(cpu_buf, pc, 0);
328}
329
1da177e4
LT
330/*
331 * This serves to avoid cpu buffer overflow, and makes sure
332 * the task mortuary progresses
333 *
334 * By using schedule_delayed_work_on and then schedule_delayed_work
335 * we guarantee this will stay on the correct cpu
336 */
c4028958 337static void wq_sync_buffer(struct work_struct *work)
1da177e4 338{
25ad2913 339 struct oprofile_cpu_buffer *b =
c4028958 340 container_of(work, struct oprofile_cpu_buffer, work.work);
1da177e4 341 if (b->cpu != smp_processor_id()) {
bd17b625 342 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
1da177e4 343 smp_processor_id(), b->cpu);
4bd9b9dc
CA
344
345 if (!cpu_online(b->cpu)) {
346 cancel_delayed_work(&b->work);
347 return;
348 }
1da177e4
LT
349 }
350 sync_buffer(b->cpu);
351
352 /* don't re-add the work if we're shutting down */
353 if (work_enabled)
354 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
355}
This page took 0.563737 seconds and 5 git commands to generate.