list_add_rcu(&counter->event_entry, &ctx->event_list);
ctx->nr_counters++;
+ if (counter->attr.inherit_stat)
+ ctx->nr_stat++;
}
/*
if (list_empty(&counter->list_entry))
return;
ctx->nr_counters--;
+ if (counter->attr.inherit_stat)
+ ctx->nr_stat--;
list_del_init(&counter->list_entry);
list_del_rcu(&counter->event_entry);
&& !ctx1->pin_count && !ctx2->pin_count;
}
+static void __perf_counter_read(void *counter);
+
+static void __perf_counter_sync_stat(struct perf_counter *counter,
+ struct perf_counter *next_counter)
+{
+ u64 value;
+
+ if (!counter->attr.inherit_stat)
+ return;
+
+ /*
+ * Update the counter value, we cannot use perf_counter_read()
+ * because we're in the middle of a context switch and have IRQs
+ * disabled, which upsets smp_call_function_single(), however
+ * we know the counter must be on the current CPU, therefore we
+ * don't need to use it.
+ */
+ switch (counter->state) {
+ case PERF_COUNTER_STATE_ACTIVE:
+ __perf_counter_read(counter);
+ break;
+
+ case PERF_COUNTER_STATE_INACTIVE:
+ update_counter_times(counter);
+ break;
+
+ default:
+ break;
+ }
+
+ /*
+ * In order to keep per-task stats reliable we need to flip the counter
+ * values when we flip the contexts.
+ */
+ value = atomic64_read(&next_counter->count);
+ value = atomic64_xchg(&counter->count, value);
+ atomic64_set(&next_counter->count, value);
+
+ swap(counter->total_time_enabled, next_counter->total_time_enabled);
+ swap(counter->total_time_running, next_counter->total_time_running);
+
+ /*
+ * Since we swizzled the values, update the user visible data too.
+ */
+ perf_counter_update_userpage(counter);
+ perf_counter_update_userpage(next_counter);
+}
+
+#define list_next_entry(pos, member) \
+ list_entry(pos->member.next, typeof(*pos), member)
+
+static void perf_counter_sync_stat(struct perf_counter_context *ctx,
+ struct perf_counter_context *next_ctx)
+{
+ struct perf_counter *counter, *next_counter;
+
+ if (!ctx->nr_stat)
+ return;
+
+ counter = list_first_entry(&ctx->event_list,
+ struct perf_counter, event_entry);
+
+ next_counter = list_first_entry(&next_ctx->event_list,
+ struct perf_counter, event_entry);
+
+ while (&counter->event_entry != &ctx->event_list &&
+ &next_counter->event_entry != &next_ctx->event_list) {
+
+ __perf_counter_sync_stat(counter, next_counter);
+
+ counter = list_next_entry(counter, event_entry);
+ next_counter = list_next_entry(counter, event_entry);
+ }
+}
+
/*
* Called from scheduler to remove the counters of the current task,
* with interrupts disabled.
ctx->task = next;
next_ctx->task = task;
do_switch = 0;
+
+ perf_counter_sync_stat(ctx, next_ctx);
}
spin_unlock(&next_ctx->lock);
spin_unlock(&ctx->lock);
perf_counter_task_sched_in(curr, cpu);
}
+/*
+ * Enable all of a task's counters that have been marked enable-on-exec.
+ * This expects task == current.
+ */
+static void perf_counter_enable_on_exec(struct task_struct *task)
+{
+ struct perf_counter_context *ctx;
+ struct perf_counter *counter;
+ unsigned long flags;
+ int enabled = 0;
+
+ local_irq_save(flags);
+ ctx = task->perf_counter_ctxp;
+ if (!ctx || !ctx->nr_counters)
+ goto out;
+
+ __perf_counter_task_sched_out(ctx);
+
+ spin_lock(&ctx->lock);
+
+ list_for_each_entry(counter, &ctx->counter_list, list_entry) {
+ if (!counter->attr.enable_on_exec)
+ continue;
+ counter->attr.enable_on_exec = 0;
+ if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
+ continue;
+ counter->state = PERF_COUNTER_STATE_INACTIVE;
+ counter->tstamp_enabled =
+ ctx->time - counter->total_time_enabled;
+ enabled = 1;
+ }
+
+ /*
+ * Unclone this context if we enabled any counter.
+ */
+ if (enabled && ctx->parent_ctx) {
+ put_ctx(ctx->parent_ctx);
+ ctx->parent_ctx = NULL;
+ }
+
+ spin_unlock(&ctx->lock);
+
+ perf_counter_task_sched_in(task, smp_processor_id());
+ out:
+ local_irq_restore(flags);
+}
+
/*
* Cross CPU call to read the hardware counter
*/
-static void __read(void *info)
+static void __perf_counter_read(void *info)
{
struct perf_counter *counter = info;
struct perf_counter_context *ctx = counter->ctx;
*/
if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
smp_call_function_single(counter->oncpu,
- __read, counter, 1);
+ __perf_counter_read, counter, 1);
} else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
update_counter_times(counter);
}
{
perf_pending_sync(counter);
- atomic_dec(&nr_counters);
- if (counter->attr.mmap)
- atomic_dec(&nr_mmap_counters);
- if (counter->attr.comm)
- atomic_dec(&nr_comm_counters);
+ if (!counter->parent) {
+ atomic_dec(&nr_counters);
+ if (counter->attr.mmap)
+ atomic_dec(&nr_mmap_counters);
+ if (counter->attr.comm)
+ atomic_dec(&nr_comm_counters);
+ }
if (counter->destroy)
counter->destroy(counter);
return 0;
}
+static int perf_counter_index(struct perf_counter *counter)
+{
+ if (counter->state != PERF_COUNTER_STATE_ACTIVE)
+ return 0;
+
+ return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
+}
+
/*
* Callers need to ensure there can be no nesting of this function, otherwise
* the seqlock logic goes bad. We can not serialize this because the arch
preempt_disable();
++userpg->lock;
barrier();
- userpg->index = counter->hw.idx;
+ userpg->index = perf_counter_index(counter);
userpg->offset = atomic64_read(&counter->count);
if (counter->state == PERF_COUNTER_STATE_ACTIVE)
userpg->offset -= atomic64_read(&counter->hw.prev_count);
+ userpg->time_enabled = counter->total_time_enabled +
+ atomic64_read(&counter->child_total_time_enabled);
+
+ userpg->time_running = counter->total_time_running +
+ atomic64_read(&counter->child_total_time_running);
+
barrier();
++userpg->lock;
preempt_enable();
static void perf_mmap_free_page(unsigned long addr)
{
- struct page *page = virt_to_page(addr);
+ struct page *page = virt_to_page((void *)addr);
page->mapping = NULL;
__free_page(page);
u32 cpu, reserved;
} cpu_entry;
- header.type = 0;
+ header.type = PERF_EVENT_SAMPLE;
header.size = sizeof(header);
- header.misc = PERF_EVENT_MISC_OVERFLOW;
+ header.misc = 0;
header.misc |= perf_misc_flags(data->regs);
if (sample_type & PERF_SAMPLE_IP) {
ip = perf_instruction_pointer(data->regs);
- header.type |= PERF_SAMPLE_IP;
header.size += sizeof(ip);
}
tid_entry.pid = perf_counter_pid(counter, current);
tid_entry.tid = perf_counter_tid(counter, current);
- header.type |= PERF_SAMPLE_TID;
header.size += sizeof(tid_entry);
}
*/
time = sched_clock();
- header.type |= PERF_SAMPLE_TIME;
header.size += sizeof(u64);
}
- if (sample_type & PERF_SAMPLE_ADDR) {
- header.type |= PERF_SAMPLE_ADDR;
+ if (sample_type & PERF_SAMPLE_ADDR)
header.size += sizeof(u64);
- }
- if (sample_type & PERF_SAMPLE_ID) {
- header.type |= PERF_SAMPLE_ID;
+ if (sample_type & PERF_SAMPLE_ID)
header.size += sizeof(u64);
- }
if (sample_type & PERF_SAMPLE_CPU) {
- header.type |= PERF_SAMPLE_CPU;
header.size += sizeof(cpu_entry);
cpu_entry.cpu = raw_smp_processor_id();
+ cpu_entry.reserved = 0;
}
- if (sample_type & PERF_SAMPLE_PERIOD) {
- header.type |= PERF_SAMPLE_PERIOD;
+ if (sample_type & PERF_SAMPLE_PERIOD)
header.size += sizeof(u64);
- }
if (sample_type & PERF_SAMPLE_GROUP) {
- header.type |= PERF_SAMPLE_GROUP;
header.size += sizeof(u64) +
counter->nr_siblings * sizeof(group_entry);
}
if (callchain) {
callchain_size = (1 + callchain->nr) * sizeof(u64);
-
- header.type |= PERF_SAMPLE_CALLCHAIN;
header.size += callchain_size;
- }
+ } else
+ header.size += sizeof(u64);
}
ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
}
}
- if (callchain)
- perf_output_copy(&handle, callchain, callchain_size);
+ if (sample_type & PERF_SAMPLE_CALLCHAIN) {
+ if (callchain)
+ perf_output_copy(&handle, callchain, callchain_size);
+ else {
+ u64 nr = 0;
+ perf_output_put(&handle, nr);
+ }
+ }
perf_output_end(&handle);
}
+/*
+ * read event
+ */
+
+struct perf_read_event {
+ struct perf_event_header header;
+
+ u32 pid;
+ u32 tid;
+ u64 value;
+ u64 format[3];
+};
+
+static void
+perf_counter_read_event(struct perf_counter *counter,
+ struct task_struct *task)
+{
+ struct perf_output_handle handle;
+ struct perf_read_event event = {
+ .header = {
+ .type = PERF_EVENT_READ,
+ .misc = 0,
+ .size = sizeof(event) - sizeof(event.format),
+ },
+ .pid = perf_counter_pid(counter, task),
+ .tid = perf_counter_tid(counter, task),
+ .value = atomic64_read(&counter->count),
+ };
+ int ret, i = 0;
+
+ if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
+ event.header.size += sizeof(u64);
+ event.format[i++] = counter->total_time_enabled;
+ }
+
+ if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
+ event.header.size += sizeof(u64);
+ event.format[i++] = counter->total_time_running;
+ }
+
+ if (counter->attr.read_format & PERF_FORMAT_ID) {
+ u64 id;
+
+ event.header.size += sizeof(u64);
+ if (counter->parent)
+ id = counter->parent->id;
+ else
+ id = counter->id;
+
+ event.format[i++] = id;
+ }
+
+ ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
+ if (ret)
+ return;
+
+ perf_output_copy(&handle, &event, event.header.size);
+ perf_output_end(&handle);
+}
+
/*
* fork tracking
*/
{
struct perf_comm_event comm_event;
+ if (task->perf_counter_ctxp)
+ perf_counter_enable_on_exec(task);
+
if (!atomic_read(&nr_comm_counters))
return;
put_cpu_var(perf_cpu_context);
}
-void
-perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
+void __perf_swcounter_event(u32 event, u64 nr, int nmi,
+ struct pt_regs *regs, u64 addr)
{
struct perf_sample_data data = {
.regs = regs,
}
#endif
+atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
+
+static void sw_perf_counter_destroy(struct perf_counter *counter)
+{
+ u64 event = counter->attr.config;
+
+ WARN_ON(counter->parent);
+
+ atomic_dec(&perf_swcounter_enabled[event]);
+}
+
static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
{
const struct pmu *pmu = NULL;
+ u64 event = counter->attr.config;
/*
* Software counters (currently) can't in general distinguish
* to be kernel events, and page faults are never hypervisor
* events.
*/
- switch (counter->attr.config) {
+ switch (event) {
case PERF_COUNT_SW_CPU_CLOCK:
pmu = &perf_ops_cpu_clock;
case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
case PERF_COUNT_SW_CONTEXT_SWITCHES:
case PERF_COUNT_SW_CPU_MIGRATIONS:
+ if (!counter->parent) {
+ atomic_inc(&perf_swcounter_enabled[event]);
+ counter->destroy = sw_perf_counter_destroy;
+ }
pmu = &perf_ops_generic;
break;
}
int cpu,
struct perf_counter_context *ctx,
struct perf_counter *group_leader,
+ struct perf_counter *parent_counter,
gfp_t gfpflags)
{
const struct pmu *pmu;
counter->ctx = ctx;
counter->oncpu = -1;
+ counter->parent = parent_counter;
+
counter->ns = get_pid_ns(current->nsproxy->pid_ns);
counter->id = atomic64_inc_return(&perf_counter_id);
counter->pmu = pmu;
- atomic_inc(&nr_counters);
- if (counter->attr.mmap)
- atomic_inc(&nr_mmap_counters);
- if (counter->attr.comm)
- atomic_inc(&nr_comm_counters);
+ if (!counter->parent) {
+ atomic_inc(&nr_counters);
+ if (counter->attr.mmap)
+ atomic_inc(&nr_mmap_counters);
+ if (counter->attr.comm)
+ atomic_inc(&nr_comm_counters);
+ }
return counter;
}
}
counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
- GFP_KERNEL);
+ NULL, GFP_KERNEL);
ret = PTR_ERR(counter);
if (IS_ERR(counter))
goto err_put_context;
child_counter = perf_counter_alloc(&parent_counter->attr,
parent_counter->cpu, child_ctx,
- group_leader, GFP_KERNEL);
+ group_leader, parent_counter,
+ GFP_KERNEL);
if (IS_ERR(child_counter))
return child_counter;
get_ctx(child_ctx);
*/
add_counter_to_ctx(child_counter, child_ctx);
- child_counter->parent = parent_counter;
- /*
- * inherit into child's child as well:
- */
- child_counter->attr.inherit = 1;
-
/*
* Get a reference to the parent filp - we will fput it
* when the child counter exits. This is safe to do because
}
static void sync_child_counter(struct perf_counter *child_counter,
- struct perf_counter *parent_counter)
+ struct task_struct *child)
{
+ struct perf_counter *parent_counter = child_counter->parent;
u64 child_val;
+ if (child_counter->attr.inherit_stat)
+ perf_counter_read_event(child_counter, child);
+
child_val = atomic64_read(&child_counter->count);
/*
static void
__perf_counter_exit_task(struct perf_counter *child_counter,
- struct perf_counter_context *child_ctx)
+ struct perf_counter_context *child_ctx,
+ struct task_struct *child)
{
struct perf_counter *parent_counter;
* counters need to be zapped - but otherwise linger.
*/
if (parent_counter) {
- sync_child_counter(child_counter, parent_counter);
+ sync_child_counter(child_counter, child);
free_counter(child_counter);
}
}
again:
list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
list_entry)
- __perf_counter_exit_task(child_counter, child_ctx);
+ __perf_counter_exit_task(child_counter, child_ctx, child);
/*
* If the last counter was a group counter, it will have appended all