4 * @remark Copyright 2002 OProfile authors
5 * @remark Read the file COPYING
7 * @author John Levon <levon@movementarian.org>
8 * @author Barry Kasindorf <barry.kasindorf@amd.com>
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().
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.
21 #include <linux/sched.h>
22 #include <linux/oprofile.h>
23 #include <linux/vmalloc.h>
24 #include <linux/errno.h>
26 #include "event_buffer.h"
27 #include "cpu_buffer.h"
28 #include "buffer_sync.h"
31 #define OP_BUFFER_FLAGS 0
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.
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.
48 struct ring_buffer
*op_ring_buffer_read
;
49 struct ring_buffer
*op_ring_buffer_write
;
50 DEFINE_PER_CPU(struct oprofile_cpu_buffer
, cpu_buffer
);
52 static void wq_sync_buffer(struct work_struct
*work
);
54 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
55 static int work_enabled
;
57 void free_cpu_buffers(void)
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
;
67 unsigned long oprofile_get_cpu_buffer_size(void)
69 return fs_cpu_buffer_size
;
72 void oprofile_cpu_buffer_inc_smpl_lost(void)
74 struct oprofile_cpu_buffer
*cpu_buf
75 = &__get_cpu_var(cpu_buffer
);
77 cpu_buf
->sample_lost_overflow
++;
80 int alloc_cpu_buffers(void)
84 unsigned long buffer_size
= fs_cpu_buffer_size
;
86 op_ring_buffer_read
= ring_buffer_alloc(buffer_size
, OP_BUFFER_FLAGS
);
87 if (!op_ring_buffer_read
)
89 op_ring_buffer_write
= ring_buffer_alloc(buffer_size
, OP_BUFFER_FLAGS
);
90 if (!op_ring_buffer_write
)
93 for_each_possible_cpu(i
) {
94 struct oprofile_cpu_buffer
*b
= &per_cpu(cpu_buffer
, i
);
97 b
->last_is_kernel
= -1;
99 b
->buffer_size
= buffer_size
;
102 b
->sample_received
= 0;
103 b
->sample_lost_overflow
= 0;
104 b
->backtrace_aborted
= 0;
105 b
->sample_invalid_eip
= 0;
107 INIT_DELAYED_WORK(&b
->work
, wq_sync_buffer
);
116 void start_cpu_work(void)
122 for_each_online_cpu(i
) {
123 struct oprofile_cpu_buffer
*b
= &per_cpu(cpu_buffer
, i
);
126 * Spread the work by 1 jiffy per cpu so they dont all
129 schedule_delayed_work_on(i
, &b
->work
, DEFAULT_TIMER_EXPIRE
+ i
);
133 void end_cpu_work(void)
139 for_each_online_cpu(i
) {
140 struct oprofile_cpu_buffer
*b
= &per_cpu(cpu_buffer
, i
);
142 cancel_delayed_work(&b
->work
);
145 flush_scheduled_work();
149 add_sample(struct oprofile_cpu_buffer
*cpu_buf
,
150 unsigned long pc
, unsigned long event
)
152 struct op_entry entry
;
155 ret
= cpu_buffer_write_entry(&entry
);
159 entry
.sample
->eip
= pc
;
160 entry
.sample
->event
= event
;
162 ret
= cpu_buffer_write_commit(&entry
);
170 add_code(struct oprofile_cpu_buffer
*buffer
, unsigned long value
)
172 return add_sample(buffer
, ESCAPE_CODE
, value
);
175 /* This must be safe from any context. It's safe writing here
176 * because of the head/tail separation of the writer and reader
179 * is_kernel is needed because on some architectures you cannot
180 * tell if you are in kernel or user space simply by looking at
181 * pc. We tag this in the buffer by generating kernel enter/exit
182 * events whenever is_kernel changes
184 static int log_sample(struct oprofile_cpu_buffer
*cpu_buf
, unsigned long pc
,
185 int is_kernel
, unsigned long event
)
187 struct task_struct
*task
;
189 cpu_buf
->sample_received
++;
191 if (pc
== ESCAPE_CODE
) {
192 cpu_buf
->sample_invalid_eip
++;
196 is_kernel
= !!is_kernel
;
200 /* notice a switch from user->kernel or vice versa */
201 if (cpu_buf
->last_is_kernel
!= is_kernel
) {
202 cpu_buf
->last_is_kernel
= is_kernel
;
203 if (add_code(cpu_buf
, is_kernel
))
207 /* notice a task switch */
208 if (cpu_buf
->last_task
!= task
) {
209 cpu_buf
->last_task
= task
;
210 if (add_code(cpu_buf
, (unsigned long)task
))
214 if (add_sample(cpu_buf
, pc
, event
))
220 cpu_buf
->sample_lost_overflow
++;
224 static int oprofile_begin_trace(struct oprofile_cpu_buffer
*cpu_buf
)
226 add_code(cpu_buf
, CPU_TRACE_BEGIN
);
227 cpu_buf
->tracing
= 1;
231 static void oprofile_end_trace(struct oprofile_cpu_buffer
*cpu_buf
)
233 cpu_buf
->tracing
= 0;
236 void oprofile_add_ext_sample(unsigned long pc
, struct pt_regs
* const regs
,
237 unsigned long event
, int is_kernel
)
239 struct oprofile_cpu_buffer
*cpu_buf
= &__get_cpu_var(cpu_buffer
);
241 if (!backtrace_depth
) {
242 log_sample(cpu_buf
, pc
, is_kernel
, event
);
246 if (!oprofile_begin_trace(cpu_buf
))
250 * if log_sample() fail we can't backtrace since we lost the
251 * source of this event
253 if (log_sample(cpu_buf
, pc
, is_kernel
, event
))
254 oprofile_ops
.backtrace(regs
, backtrace_depth
);
255 oprofile_end_trace(cpu_buf
);
258 void oprofile_add_sample(struct pt_regs
* const regs
, unsigned long event
)
260 int is_kernel
= !user_mode(regs
);
261 unsigned long pc
= profile_pc(regs
);
263 oprofile_add_ext_sample(pc
, regs
, event
, is_kernel
);
266 #ifdef CONFIG_OPROFILE_IBS
268 #define MAX_IBS_SAMPLE_SIZE 14
270 void oprofile_add_ibs_sample(struct pt_regs
* const regs
,
271 unsigned int * const ibs_sample
, int ibs_code
)
273 int is_kernel
= !user_mode(regs
);
274 struct oprofile_cpu_buffer
*cpu_buf
= &__get_cpu_var(cpu_buffer
);
275 struct task_struct
*task
;
278 cpu_buf
->sample_received
++;
280 /* notice a switch from user->kernel or vice versa */
281 if (cpu_buf
->last_is_kernel
!= is_kernel
) {
282 if (add_code(cpu_buf
, is_kernel
))
284 cpu_buf
->last_is_kernel
= is_kernel
;
287 /* notice a task switch */
290 if (cpu_buf
->last_task
!= task
) {
291 if (add_code(cpu_buf
, (unsigned long)task
))
293 cpu_buf
->last_task
= task
;
297 fail
= fail
|| add_code(cpu_buf
, ibs_code
);
298 fail
= fail
|| add_sample(cpu_buf
, ibs_sample
[0], ibs_sample
[1]);
299 fail
= fail
|| add_sample(cpu_buf
, ibs_sample
[2], ibs_sample
[3]);
300 fail
= fail
|| add_sample(cpu_buf
, ibs_sample
[4], ibs_sample
[5]);
302 if (ibs_code
== IBS_OP_BEGIN
) {
303 fail
= fail
|| add_sample(cpu_buf
, ibs_sample
[6], ibs_sample
[7]);
304 fail
= fail
|| add_sample(cpu_buf
, ibs_sample
[8], ibs_sample
[9]);
305 fail
= fail
|| add_sample(cpu_buf
, ibs_sample
[10], ibs_sample
[11]);
312 oprofile_ops
.backtrace(regs
, backtrace_depth
);
317 cpu_buf
->sample_lost_overflow
++;
323 void oprofile_add_pc(unsigned long pc
, int is_kernel
, unsigned long event
)
325 struct oprofile_cpu_buffer
*cpu_buf
= &__get_cpu_var(cpu_buffer
);
326 log_sample(cpu_buf
, pc
, is_kernel
, event
);
329 void oprofile_add_trace(unsigned long pc
)
331 struct oprofile_cpu_buffer
*cpu_buf
= &__get_cpu_var(cpu_buffer
);
333 if (!cpu_buf
->tracing
)
337 * broken frame can give an eip with the same value as an
338 * escape code, abort the trace if we get it
340 if (pc
== ESCAPE_CODE
)
343 if (add_sample(cpu_buf
, pc
, 0))
348 cpu_buf
->tracing
= 0;
349 cpu_buf
->backtrace_aborted
++;
354 * This serves to avoid cpu buffer overflow, and makes sure
355 * the task mortuary progresses
357 * By using schedule_delayed_work_on and then schedule_delayed_work
358 * we guarantee this will stay on the correct cpu
360 static void wq_sync_buffer(struct work_struct
*work
)
362 struct oprofile_cpu_buffer
*b
=
363 container_of(work
, struct oprofile_cpu_buffer
, work
.work
);
364 if (b
->cpu
!= smp_processor_id()) {
365 printk(KERN_DEBUG
"WQ on CPU%d, prefer CPU%d\n",
366 smp_processor_id(), b
->cpu
);
368 if (!cpu_online(b
->cpu
)) {
369 cancel_delayed_work(&b
->work
);
375 /* don't re-add the work if we're shutting down */
377 schedule_delayed_work(&b
->work
, DEFAULT_TIMER_EXPIRE
);
This page took 0.038744 seconds and 5 git commands to generate.