*
* @author John Levon <levon@movementarian.org>
*
+ * Modified by Aravind Menon for Xen
+ * These modifications are:
+ * Copyright (C) 2005 Hewlett-Packard Co.
+ *
* Each CPU has a local buffer that stores PC value/event
* pairs. We also log context switches when we notice them.
* Eventually each CPU's buffer is processed into the global
- * event buffer by sync_cpu_buffers().
+ * event buffer by sync_buffer().
*
* We use a local buffer for two reasons: an NMI or similar
* interrupt cannot synchronise, and high sampling rates
*/
#include <linux/sched.h>
+#include <linux/oprofile.h>
#include <linux/vmalloc.h>
#include <linux/errno.h>
+#include "event_buffer.h"
#include "cpu_buffer.h"
+#include "buffer_sync.h"
#include "oprof.h"
struct oprofile_cpu_buffer cpu_buffer[NR_CPUS] __cacheline_aligned;
-static void __free_cpu_buffers(int num)
+static void wq_sync_buffer(struct work_struct *work);
+
+#define DEFAULT_TIMER_EXPIRE (HZ / 10)
+static int work_enabled;
+
+static int32_t current_domain = COORDINATOR_DOMAIN;
+
+void free_cpu_buffers(void)
{
int i;
- for (i=0; i < num; ++i) {
- struct oprofile_cpu_buffer * b = &cpu_buffer[i];
-
- if (!cpu_possible(i))
- continue;
-
- vfree(b->buffer);
- }
+ for_each_online_cpu(i)
+ vfree(cpu_buffer[i].buffer);
}
-
-
+
int alloc_cpu_buffers(void)
{
int i;
unsigned long buffer_size = fs_cpu_buffer_size;
- for (i=0; i < NR_CPUS; ++i) {
+ for_each_online_cpu(i) {
struct oprofile_cpu_buffer * b = &cpu_buffer[i];
- if (!cpu_possible(i))
- continue;
-
- b->buffer = vmalloc(sizeof(struct op_sample) * buffer_size);
+ b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
+ cpu_to_node(i));
if (!b->buffer)
goto fail;
- b->last_task = 0;
- b->last_is_kernel = -1;
+ b->last_task = NULL;
+ b->last_cpu_mode = -1;
+ b->tracing = 0;
b->buffer_size = buffer_size;
b->tail_pos = 0;
b->head_pos = 0;
b->sample_received = 0;
b->sample_lost_overflow = 0;
- b->sample_lost_task_exit = 0;
+ b->cpu = i;
+ INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
}
return 0;
+
fail:
- __free_cpu_buffers(i);
+ free_cpu_buffers();
return -ENOMEM;
}
-
-void free_cpu_buffers(void)
+void start_cpu_work(void)
{
- __free_cpu_buffers(NR_CPUS);
+ int i;
+
+ work_enabled = 1;
+
+ for_each_online_cpu(i) {
+ struct oprofile_cpu_buffer * b = &cpu_buffer[i];
+
+ /*
+ * Spread the work by 1 jiffy per cpu so they dont all
+ * fire at once.
+ */
+ schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
+ }
+}
+
+void end_cpu_work(void)
+{
+ int i;
+
+ work_enabled = 0;
+
+ for_each_online_cpu(i) {
+ struct oprofile_cpu_buffer * b = &cpu_buffer[i];
+
+ cancel_delayed_work(&b->work);
+ }
+
+ flush_scheduled_work();
}
+/* Resets the cpu buffer to a sane state. */
+void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
+{
+ /* reset these to invalid values; the next sample
+ * collected will populate the buffer with proper
+ * values to initialize the buffer
+ */
+ cpu_buf->last_cpu_mode = -1;
+ cpu_buf->last_task = NULL;
+}
/* compute number of available slots in cpu_buffer queue */
static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
return tail + (b->buffer_size - head) - 1;
}
-
static void increment_head(struct oprofile_cpu_buffer * b)
{
unsigned long new_head = b->head_pos + 1;
* increment is visible */
wmb();
- if (new_head < (b->buffer_size))
+ if (new_head < b->buffer_size)
b->head_pos = new_head;
else
b->head_pos = 0;
}
+static inline void
+add_sample(struct oprofile_cpu_buffer * cpu_buf,
+ unsigned long pc, unsigned long event)
+{
+ struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
+ entry->eip = pc;
+ entry->event = event;
+ increment_head(cpu_buf);
+}
+
+static inline void
+add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
+{
+ add_sample(buffer, ESCAPE_CODE, value);
+}
/* This must be safe from any context. It's safe writing here
* because of the head/tail separation of the writer and reader
* of the CPU buffer.
*
- * is_kernel is needed because on some architectures you cannot
+ * cpu_mode is needed because on some architectures you cannot
* tell if you are in kernel or user space simply by looking at
- * eip. We tag this in the buffer by generating kernel enter/exit
- * events whenever is_kernel changes
+ * pc. We tag this in the buffer by generating kernel/user (and xen)
+ * enter events whenever cpu_mode changes
*/
-void oprofile_add_sample(unsigned long eip, unsigned int is_kernel,
- unsigned long event, int cpu)
+static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
+ int cpu_mode, unsigned long event)
{
- struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[cpu];
struct task_struct * task;
- is_kernel = !!is_kernel;
-
cpu_buf->sample_received++;
-
if (nr_available_slots(cpu_buf) < 3) {
cpu_buf->sample_lost_overflow++;
- return;
+ return 0;
}
+ WARN_ON(cpu_mode > CPU_MODE_XEN);
+
task = current;
/* notice a switch from user->kernel or vice versa */
- if (cpu_buf->last_is_kernel != is_kernel) {
- cpu_buf->last_is_kernel = is_kernel;
- cpu_buf->buffer[cpu_buf->head_pos].eip = ~0UL;
- cpu_buf->buffer[cpu_buf->head_pos].event = is_kernel;
- increment_head(cpu_buf);
+ if (cpu_buf->last_cpu_mode != cpu_mode) {
+ cpu_buf->last_cpu_mode = cpu_mode;
+ add_code(cpu_buf, cpu_mode);
}
-
+
/* notice a task switch */
- if (cpu_buf->last_task != task) {
+ /* if not processing other domain samples */
+ if ((cpu_buf->last_task != task) &&
+ (current_domain == COORDINATOR_DOMAIN)) {
cpu_buf->last_task = task;
- if (!(task->flags & PF_EXITING)) {
- cpu_buf->buffer[cpu_buf->head_pos].eip = ~0UL;
- cpu_buf->buffer[cpu_buf->head_pos].event = (unsigned long)task;
- increment_head(cpu_buf);
- }
+ add_code(cpu_buf, (unsigned long)task);
}
- /* If the task is exiting it's not safe to take a sample
- * as the task_struct is about to be freed. We can't just
- * notify at release_task() time because of CLONE_DETACHED
- * tasks that release_task() themselves.
- */
- if (task->flags & PF_EXITING) {
- cpu_buf->sample_lost_task_exit++;
+ add_sample(cpu_buf, pc, event);
+ return 1;
+}
+
+static int oprofile_begin_trace(struct oprofile_cpu_buffer * cpu_buf)
+{
+ if (nr_available_slots(cpu_buf) < 4) {
+ cpu_buf->sample_lost_overflow++;
+ return 0;
+ }
+
+ add_code(cpu_buf, CPU_TRACE_BEGIN);
+ cpu_buf->tracing = 1;
+ return 1;
+}
+
+static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
+{
+ cpu_buf->tracing = 0;
+}
+
+void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
+ unsigned long event, int is_kernel)
+{
+ struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
+
+ if (!backtrace_depth) {
+ log_sample(cpu_buf, pc, is_kernel, event);
return;
}
-
- cpu_buf->buffer[cpu_buf->head_pos].eip = eip;
- cpu_buf->buffer[cpu_buf->head_pos].event = event;
- increment_head(cpu_buf);
+
+ if (!oprofile_begin_trace(cpu_buf))
+ return;
+
+ /* if log_sample() fail we can't backtrace since we lost the source
+ * of this event */
+ if (log_sample(cpu_buf, pc, is_kernel, event))
+ oprofile_ops.backtrace(regs, backtrace_depth);
+ oprofile_end_trace(cpu_buf);
}
+void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
+{
+ int is_kernel = !user_mode(regs);
+ unsigned long pc = profile_pc(regs);
-/* Resets the cpu buffer to a sane state. */
-void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
+ oprofile_add_ext_sample(pc, regs, event, is_kernel);
+}
+
+void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
{
- /* reset these to invalid values; the next sample
- * collected will populate the buffer with proper
- * values to initialize the buffer
- */
- cpu_buf->last_is_kernel = -1;
- cpu_buf->last_task = 0;
+ struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
+ log_sample(cpu_buf, pc, is_kernel, event);
+}
+
+void oprofile_add_trace(unsigned long pc)
+{
+ struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
+
+ if (!cpu_buf->tracing)
+ return;
+
+ if (nr_available_slots(cpu_buf) < 1) {
+ cpu_buf->tracing = 0;
+ cpu_buf->sample_lost_overflow++;
+ return;
+ }
+
+ /* broken frame can give an eip with the same value as an escape code,
+ * abort the trace if we get it */
+ if (pc == ESCAPE_CODE) {
+ cpu_buf->tracing = 0;
+ cpu_buf->backtrace_aborted++;
+ return;
+ }
+
+ add_sample(cpu_buf, pc, 0);
+}
+
+int oprofile_add_domain_switch(int32_t domain_id)
+{
+ struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
+
+ /* should have space for switching into and out of domain
+ (2 slots each) plus one sample and one cpu mode switch */
+ if (((nr_available_slots(cpu_buf) < 6) &&
+ (domain_id != COORDINATOR_DOMAIN)) ||
+ (nr_available_slots(cpu_buf) < 2))
+ return 0;
+
+ add_code(cpu_buf, CPU_DOMAIN_SWITCH);
+ add_sample(cpu_buf, domain_id, 0);
+
+ current_domain = domain_id;
+
+ return 1;
+}
+
+/*
+ * This serves to avoid cpu buffer overflow, and makes sure
+ * the task mortuary progresses
+ *
+ * By using schedule_delayed_work_on and then schedule_delayed_work
+ * we guarantee this will stay on the correct cpu
+ */
+static void wq_sync_buffer(struct work_struct *work)
+{
+ struct oprofile_cpu_buffer * b =
+ container_of(work, struct oprofile_cpu_buffer, work.work);
+ if (b->cpu != smp_processor_id()) {
+ printk("WQ on CPU%d, prefer CPU%d\n",
+ smp_processor_id(), b->cpu);
+ }
+ sync_buffer(b->cpu);
+
+ /* don't re-add the work if we're shutting down */
+ if (work_enabled)
+ schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
}
git://git.onelab.eu / linux-2.6.git / blobdiff