/*
* Performance counter core code
*
- * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
- * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
- *
+ * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
+ * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
+ * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
+ * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
*
* For licensing details see kernel-base/COPYING
*/
static int perf_reserved_percpu __read_mostly;
static int perf_overcommit __read_mostly = 1;
+static atomic_t nr_counters __read_mostly;
+static atomic_t nr_mmap_tracking __read_mostly;
+static atomic_t nr_munmap_tracking __read_mostly;
+static atomic_t nr_comm_tracking __read_mostly;
+
+int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
+int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
+
/*
- * Mutex for (sysadmin-configurable) counter reservations:
+ * Lock for (sysadmin-configurable) counter reservations:
*/
-static DEFINE_MUTEX(perf_resource_mutex);
+static DEFINE_SPINLOCK(perf_resource_lock);
/*
* Architecture provided APIs - weak aliases:
*/
-extern __weak const struct hw_perf_counter_ops *
-hw_perf_counter_init(struct perf_counter *counter)
+extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
{
return NULL;
}
-u64 __weak hw_perf_save_disable(void) { return 0; }
-void __weak hw_perf_restore(u64 ctrl) { barrier(); }
+void __weak hw_perf_disable(void) { barrier(); }
+void __weak hw_perf_enable(void) { barrier(); }
+
void __weak hw_perf_counter_setup(int cpu) { barrier(); }
int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
struct perf_cpu_context *cpuctx,
void __weak perf_counter_print_debug(void) { }
+static DEFINE_PER_CPU(int, disable_count);
+
+void __perf_disable(void)
+{
+ __get_cpu_var(disable_count)++;
+}
+
+bool __perf_enable(void)
+{
+ return !--__get_cpu_var(disable_count);
+}
+
+void perf_disable(void)
+{
+ __perf_disable();
+ hw_perf_disable();
+}
+
+void perf_enable(void)
+{
+ if (__perf_enable())
+ hw_perf_enable();
+}
+
+static void get_ctx(struct perf_counter_context *ctx)
+{
+ atomic_inc(&ctx->refcount);
+}
+
+static void put_ctx(struct perf_counter_context *ctx)
+{
+ if (atomic_dec_and_test(&ctx->refcount)) {
+ if (ctx->parent_ctx)
+ put_ctx(ctx->parent_ctx);
+ kfree(ctx);
+ }
+}
+
+/*
+ * Add a counter from the lists for its context.
+ * Must be called with ctx->mutex and ctx->lock held.
+ */
static void
list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
{
* add it straight to the context's counter list, or to the group
* leader's sibling list:
*/
- if (counter->group_leader == counter)
+ if (group_leader == counter)
list_add_tail(&counter->list_entry, &ctx->counter_list);
else {
list_add_tail(&counter->list_entry, &group_leader->sibling_list);
}
list_add_rcu(&counter->event_entry, &ctx->event_list);
+ ctx->nr_counters++;
}
+/*
+ * Remove a counter from the lists for its context.
+ * Must be called with ctx->mutex and ctx->lock held.
+ */
static void
list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
{
struct perf_counter *sibling, *tmp;
+ if (list_empty(&counter->list_entry))
+ return;
+ ctx->nr_counters--;
+
list_del_init(&counter->list_entry);
list_del_rcu(&counter->event_entry);
counter->state = PERF_COUNTER_STATE_INACTIVE;
counter->tstamp_stopped = ctx->time;
- counter->hw_ops->disable(counter);
+ counter->pmu->disable(counter);
counter->oncpu = -1;
if (!is_software_counter(counter))
cpuctx->exclusive = 0;
}
+/*
+ * Mark this context as not being a clone of another.
+ * Called when counters are added to or removed from this context.
+ * We also increment our generation number so that anything that
+ * was cloned from this context before this will not match anything
+ * cloned from this context after this.
+ */
+static void unclone_ctx(struct perf_counter_context *ctx)
+{
+ ++ctx->generation;
+ if (!ctx->parent_ctx)
+ return;
+ put_ctx(ctx->parent_ctx);
+ ctx->parent_ctx = NULL;
+}
+
/*
* Cross CPU call to remove a performance counter
*
struct perf_counter *counter = info;
struct perf_counter_context *ctx = counter->ctx;
unsigned long flags;
- u64 perf_flags;
/*
* If this is a task context, we need to check whether it is
return;
spin_lock_irqsave(&ctx->lock, flags);
-
- counter_sched_out(counter, cpuctx, ctx);
-
- counter->task = NULL;
- ctx->nr_counters--;
-
/*
* Protect the list operation against NMI by disabling the
- * counters on a global level. NOP for non NMI based counters.
+ * counters on a global level.
*/
- perf_flags = hw_perf_save_disable();
+ perf_disable();
+
+ counter_sched_out(counter, cpuctx, ctx);
+
list_del_counter(counter, ctx);
- hw_perf_restore(perf_flags);
if (!ctx->task) {
/*
perf_max_counters - perf_reserved_percpu);
}
+ perf_enable();
spin_unlock_irqrestore(&ctx->lock, flags);
}
/*
* Remove the counter from a task's (or a CPU's) list of counters.
*
- * Must be called with counter->mutex and ctx->mutex held.
+ * Must be called with ctx->mutex held.
*
* CPU counters are removed with a smp call. For task counters we only
* call when the task is on a CPU.
struct perf_counter_context *ctx = counter->ctx;
struct task_struct *task = ctx->task;
+ unclone_ctx(ctx);
if (!task) {
/*
* Per cpu counters are removed via an smp call and
* succeed.
*/
if (!list_empty(&counter->list_entry)) {
- ctx->nr_counters--;
list_del_counter(counter, ctx);
- counter->task = NULL;
}
spin_unlock_irq(&ctx->lock);
}
spin_unlock_irq(&ctx->lock);
}
-/*
- * Disable a counter and all its children.
- */
-static void perf_counter_disable_family(struct perf_counter *counter)
-{
- struct perf_counter *child;
-
- perf_counter_disable(counter);
-
- /*
- * Lock the mutex to protect the list of children
- */
- mutex_lock(&counter->mutex);
- list_for_each_entry(child, &counter->child_list, child_list)
- perf_counter_disable(child);
- mutex_unlock(&counter->mutex);
-}
-
static int
counter_sched_in(struct perf_counter *counter,
struct perf_cpu_context *cpuctx,
*/
smp_wmb();
- if (counter->hw_ops->enable(counter)) {
+ if (counter->pmu->enable(counter)) {
counter->state = PERF_COUNTER_STATE_INACTIVE;
counter->oncpu = -1;
return -EAGAIN;
return 0;
}
+static int
+group_sched_in(struct perf_counter *group_counter,
+ struct perf_cpu_context *cpuctx,
+ struct perf_counter_context *ctx,
+ int cpu)
+{
+ struct perf_counter *counter, *partial_group;
+ int ret;
+
+ if (group_counter->state == PERF_COUNTER_STATE_OFF)
+ return 0;
+
+ ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
+ if (ret)
+ return ret < 0 ? ret : 0;
+
+ group_counter->prev_state = group_counter->state;
+ if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
+ return -EAGAIN;
+
+ /*
+ * Schedule in siblings as one group (if any):
+ */
+ list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
+ counter->prev_state = counter->state;
+ if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
+ partial_group = counter;
+ goto group_error;
+ }
+ }
+
+ return 0;
+
+group_error:
+ /*
+ * Groups can be scheduled in as one unit only, so undo any
+ * partial group before returning:
+ */
+ list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
+ if (counter == partial_group)
+ break;
+ counter_sched_out(counter, cpuctx, ctx);
+ }
+ counter_sched_out(group_counter, cpuctx, ctx);
+
+ return -EAGAIN;
+}
+
/*
* Return 1 for a group consisting entirely of software counters,
* 0 if the group contains any hardware counters.
struct perf_counter_context *ctx)
{
list_add_counter(counter, ctx);
- ctx->nr_counters++;
counter->prev_state = PERF_COUNTER_STATE_OFF;
counter->tstamp_enabled = ctx->time;
counter->tstamp_running = ctx->time;
/*
* Cross CPU call to install and enable a performance counter
+ *
+ * Must be called with ctx->mutex held
*/
static void __perf_install_in_context(void *info)
{
struct perf_counter *leader = counter->group_leader;
int cpu = smp_processor_id();
unsigned long flags;
- u64 perf_flags;
int err;
/*
* If this is a task context, we need to check whether it is
* the current task context of this cpu. If not it has been
* scheduled out before the smp call arrived.
+ * Or possibly this is the right context but it isn't
+ * on this cpu because it had no counters.
*/
- if (ctx->task && cpuctx->task_ctx != ctx)
- return;
+ if (ctx->task && cpuctx->task_ctx != ctx) {
+ if (cpuctx->task_ctx || ctx->task != current)
+ return;
+ cpuctx->task_ctx = ctx;
+ }
spin_lock_irqsave(&ctx->lock, flags);
+ ctx->is_active = 1;
update_context_time(ctx);
/*
* Protect the list operation against NMI by disabling the
* counters on a global level. NOP for non NMI based counters.
*/
- perf_flags = hw_perf_save_disable();
+ perf_disable();
add_counter_to_ctx(counter, ctx);
cpuctx->max_pertask--;
unlock:
- hw_perf_restore(perf_flags);
+ perf_enable();
spin_unlock_irqrestore(&ctx->lock, flags);
}
return;
}
- counter->task = task;
retry:
task_oncpu_function_call(task, __perf_install_in_context,
counter);
* If this is a per-task counter, need to check whether this
* counter's task is the current task on this cpu.
*/
- if (ctx->task && cpuctx->task_ctx != ctx)
- return;
+ if (ctx->task && cpuctx->task_ctx != ctx) {
+ if (cpuctx->task_ctx || ctx->task != current)
+ return;
+ cpuctx->task_ctx = ctx;
+ }
spin_lock_irqsave(&ctx->lock, flags);
+ ctx->is_active = 1;
update_context_time(ctx);
counter->prev_state = counter->state;
if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
goto unlock;
- if (!group_can_go_on(counter, cpuctx, 1))
+ if (!group_can_go_on(counter, cpuctx, 1)) {
err = -EEXIST;
- else
- err = counter_sched_in(counter, cpuctx, ctx,
- smp_processor_id());
+ } else {
+ perf_disable();
+ if (counter == leader)
+ err = group_sched_in(counter, cpuctx, ctx,
+ smp_processor_id());
+ else
+ err = counter_sched_in(counter, cpuctx, ctx,
+ smp_processor_id());
+ perf_enable();
+ }
if (err) {
/*
spin_unlock_irq(&ctx->lock);
}
-static void perf_counter_refresh(struct perf_counter *counter, int refresh)
+static int perf_counter_refresh(struct perf_counter *counter, int refresh)
{
- atomic_add(refresh, &counter->event_limit);
- perf_counter_enable(counter);
-}
-
-/*
- * Enable a counter and all its children.
- */
-static void perf_counter_enable_family(struct perf_counter *counter)
-{
- struct perf_counter *child;
+ /*
+ * not supported on inherited counters
+ */
+ if (counter->hw_event.inherit)
+ return -EINVAL;
+ atomic_add(refresh, &counter->event_limit);
perf_counter_enable(counter);
- /*
- * Lock the mutex to protect the list of children
- */
- mutex_lock(&counter->mutex);
- list_for_each_entry(child, &counter->child_list, child_list)
- perf_counter_enable(child);
- mutex_unlock(&counter->mutex);
+ return 0;
}
void __perf_counter_sched_out(struct perf_counter_context *ctx,
struct perf_cpu_context *cpuctx)
{
struct perf_counter *counter;
- u64 flags;
spin_lock(&ctx->lock);
ctx->is_active = 0;
goto out;
update_context_time(ctx);
- flags = hw_perf_save_disable();
+ perf_disable();
if (ctx->nr_active) {
- list_for_each_entry(counter, &ctx->counter_list, list_entry)
- group_sched_out(counter, cpuctx, ctx);
+ list_for_each_entry(counter, &ctx->counter_list, list_entry) {
+ if (counter != counter->group_leader)
+ counter_sched_out(counter, cpuctx, ctx);
+ else
+ group_sched_out(counter, cpuctx, ctx);
+ }
}
- hw_perf_restore(flags);
+ perf_enable();
out:
spin_unlock(&ctx->lock);
}
+/*
+ * Test whether two contexts are equivalent, i.e. whether they
+ * have both been cloned from the same version of the same context
+ * and they both have the same number of enabled counters.
+ * If the number of enabled counters is the same, then the set
+ * of enabled counters should be the same, because these are both
+ * inherited contexts, therefore we can't access individual counters
+ * in them directly with an fd; we can only enable/disable all
+ * counters via prctl, or enable/disable all counters in a family
+ * via ioctl, which will have the same effect on both contexts.
+ */
+static int context_equiv(struct perf_counter_context *ctx1,
+ struct perf_counter_context *ctx2)
+{
+ return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
+ && ctx1->parent_gen == ctx2->parent_gen;
+}
+
/*
* Called from scheduler to remove the counters of the current task,
* with interrupts disabled.
* accessing the counter control register. If a NMI hits, then it will
* not restart the counter.
*/
-void perf_counter_task_sched_out(struct task_struct *task, int cpu)
+void perf_counter_task_sched_out(struct task_struct *task,
+ struct task_struct *next, int cpu)
{
struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
- struct perf_counter_context *ctx = &task->perf_counter_ctx;
+ struct perf_counter_context *ctx = task->perf_counter_ctxp;
+ struct perf_counter_context *next_ctx;
struct pt_regs *regs;
- if (likely(!cpuctx->task_ctx))
+ if (likely(!ctx || !cpuctx->task_ctx))
return;
update_context_time(ctx);
regs = task_pt_regs(task);
- perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs);
+ perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
+
+ next_ctx = next->perf_counter_ctxp;
+ if (next_ctx && context_equiv(ctx, next_ctx)) {
+ task->perf_counter_ctxp = next_ctx;
+ next->perf_counter_ctxp = ctx;
+ ctx->task = next;
+ next_ctx->task = task;
+ return;
+ }
+
__perf_counter_sched_out(ctx, cpuctx);
cpuctx->task_ctx = NULL;
}
-static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
+static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
{
- __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
+ struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
+
+ if (!cpuctx->task_ctx)
+ return;
+ __perf_counter_sched_out(ctx, cpuctx);
+ cpuctx->task_ctx = NULL;
}
-static int
-group_sched_in(struct perf_counter *group_counter,
- struct perf_cpu_context *cpuctx,
- struct perf_counter_context *ctx,
- int cpu)
+static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
{
- struct perf_counter *counter, *partial_group;
- int ret;
-
- if (group_counter->state == PERF_COUNTER_STATE_OFF)
- return 0;
-
- ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
- if (ret)
- return ret < 0 ? ret : 0;
-
- group_counter->prev_state = group_counter->state;
- if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
- return -EAGAIN;
-
- /*
- * Schedule in siblings as one group (if any):
- */
- list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
- counter->prev_state = counter->state;
- if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
- partial_group = counter;
- goto group_error;
- }
- }
-
- return 0;
-
-group_error:
- /*
- * Groups can be scheduled in as one unit only, so undo any
- * partial group before returning:
- */
- list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
- if (counter == partial_group)
- break;
- counter_sched_out(counter, cpuctx, ctx);
- }
- counter_sched_out(group_counter, cpuctx, ctx);
-
- return -EAGAIN;
+ __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
}
static void
struct perf_cpu_context *cpuctx, int cpu)
{
struct perf_counter *counter;
- u64 flags;
int can_add_hw = 1;
spin_lock(&ctx->lock);
ctx->timestamp = perf_clock();
- flags = hw_perf_save_disable();
+ perf_disable();
/*
* First go through the list and put on any pinned groups
if (counter->cpu != -1 && counter->cpu != cpu)
continue;
- if (group_can_go_on(counter, cpuctx, 1))
- group_sched_in(counter, cpuctx, ctx, cpu);
+ if (counter != counter->group_leader)
+ counter_sched_in(counter, cpuctx, ctx, cpu);
+ else {
+ if (group_can_go_on(counter, cpuctx, 1))
+ group_sched_in(counter, cpuctx, ctx, cpu);
+ }
/*
* If this pinned group hasn't been scheduled,
if (counter->cpu != -1 && counter->cpu != cpu)
continue;
- if (group_can_go_on(counter, cpuctx, can_add_hw)) {
- if (group_sched_in(counter, cpuctx, ctx, cpu))
+ if (counter != counter->group_leader) {
+ if (counter_sched_in(counter, cpuctx, ctx, cpu))
can_add_hw = 0;
+ } else {
+ if (group_can_go_on(counter, cpuctx, can_add_hw)) {
+ if (group_sched_in(counter, cpuctx, ctx, cpu))
+ can_add_hw = 0;
+ }
}
}
- hw_perf_restore(flags);
+ perf_enable();
out:
spin_unlock(&ctx->lock);
}
void perf_counter_task_sched_in(struct task_struct *task, int cpu)
{
struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
- struct perf_counter_context *ctx = &task->perf_counter_ctx;
+ struct perf_counter_context *ctx = task->perf_counter_ctxp;
+ if (likely(!ctx))
+ return;
+ if (cpuctx->task_ctx == ctx)
+ return;
__perf_counter_sched_in(ctx, cpuctx, cpu);
cpuctx->task_ctx = ctx;
}
__perf_counter_sched_in(ctx, cpuctx, cpu);
}
-int perf_counter_task_disable(void)
+int perf_counter_task_enable(void)
{
- struct task_struct *curr = current;
- struct perf_counter_context *ctx = &curr->perf_counter_ctx;
struct perf_counter *counter;
- unsigned long flags;
- u64 perf_flags;
- int cpu;
-
- if (likely(!ctx->nr_counters))
- return 0;
-
- local_irq_save(flags);
- cpu = smp_processor_id();
-
- perf_counter_task_sched_out(curr, cpu);
-
- spin_lock(&ctx->lock);
-
- /*
- * Disable all the counters:
- */
- perf_flags = hw_perf_save_disable();
- list_for_each_entry(counter, &ctx->counter_list, list_entry) {
- if (counter->state != PERF_COUNTER_STATE_ERROR) {
- update_group_times(counter);
- counter->state = PERF_COUNTER_STATE_OFF;
- }
- }
-
- hw_perf_restore(perf_flags);
-
- spin_unlock_irqrestore(&ctx->lock, flags);
+ mutex_lock(¤t->perf_counter_mutex);
+ list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
+ perf_counter_enable(counter);
+ mutex_unlock(¤t->perf_counter_mutex);
return 0;
}
-int perf_counter_task_enable(void)
+int perf_counter_task_disable(void)
{
- struct task_struct *curr = current;
- struct perf_counter_context *ctx = &curr->perf_counter_ctx;
struct perf_counter *counter;
- unsigned long flags;
- u64 perf_flags;
- int cpu;
- if (likely(!ctx->nr_counters))
- return 0;
-
- local_irq_save(flags);
- cpu = smp_processor_id();
+ mutex_lock(¤t->perf_counter_mutex);
+ list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
+ perf_counter_disable(counter);
+ mutex_unlock(¤t->perf_counter_mutex);
- perf_counter_task_sched_out(curr, cpu);
+ return 0;
+}
- spin_lock(&ctx->lock);
+static void perf_log_period(struct perf_counter *counter, u64 period);
- /*
- * Disable all the counters:
- */
- perf_flags = hw_perf_save_disable();
+static void perf_adjust_freq(struct perf_counter_context *ctx)
+{
+ struct perf_counter *counter;
+ u64 irq_period;
+ u64 events, period;
+ s64 delta;
+ spin_lock(&ctx->lock);
list_for_each_entry(counter, &ctx->counter_list, list_entry) {
- if (counter->state > PERF_COUNTER_STATE_OFF)
+ if (counter->state != PERF_COUNTER_STATE_ACTIVE)
continue;
- counter->state = PERF_COUNTER_STATE_INACTIVE;
- counter->tstamp_enabled =
- ctx->time - counter->total_time_enabled;
- counter->hw_event.disabled = 0;
- }
- hw_perf_restore(perf_flags);
- spin_unlock(&ctx->lock);
+ if (!counter->hw_event.freq || !counter->hw_event.irq_freq)
+ continue;
- perf_counter_task_sched_in(curr, cpu);
+ events = HZ * counter->hw.interrupts * counter->hw.irq_period;
+ period = div64_u64(events, counter->hw_event.irq_freq);
- local_irq_restore(flags);
+ delta = (s64)(1 + period - counter->hw.irq_period);
+ delta >>= 1;
- return 0;
+ irq_period = counter->hw.irq_period + delta;
+
+ if (!irq_period)
+ irq_period = 1;
+
+ perf_log_period(counter, irq_period);
+
+ counter->hw.irq_period = irq_period;
+ counter->hw.interrupts = 0;
+ }
+ spin_unlock(&ctx->lock);
}
/*
static void rotate_ctx(struct perf_counter_context *ctx)
{
struct perf_counter *counter;
- u64 perf_flags;
if (!ctx->nr_counters)
return;
/*
* Rotate the first entry last (works just fine for group counters too):
*/
- perf_flags = hw_perf_save_disable();
+ perf_disable();
list_for_each_entry(counter, &ctx->counter_list, list_entry) {
list_move_tail(&counter->list_entry, &ctx->counter_list);
break;
}
- hw_perf_restore(perf_flags);
+ perf_enable();
spin_unlock(&ctx->lock);
}
void perf_counter_task_tick(struct task_struct *curr, int cpu)
{
- struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
- struct perf_counter_context *ctx = &curr->perf_counter_ctx;
- const int rotate_percpu = 0;
+ struct perf_cpu_context *cpuctx;
+ struct perf_counter_context *ctx;
+
+ if (!atomic_read(&nr_counters))
+ return;
+
+ cpuctx = &per_cpu(perf_cpu_context, cpu);
+ ctx = curr->perf_counter_ctxp;
- if (rotate_percpu)
- perf_counter_cpu_sched_out(cpuctx);
- perf_counter_task_sched_out(curr, cpu);
+ perf_adjust_freq(&cpuctx->ctx);
+ if (ctx)
+ perf_adjust_freq(ctx);
- if (rotate_percpu)
- rotate_ctx(&cpuctx->ctx);
- rotate_ctx(ctx);
+ perf_counter_cpu_sched_out(cpuctx);
+ if (ctx)
+ __perf_counter_task_sched_out(ctx);
- if (rotate_percpu)
- perf_counter_cpu_sched_in(cpuctx, cpu);
- perf_counter_task_sched_in(curr, cpu);
+ rotate_ctx(&cpuctx->ctx);
+ if (ctx)
+ rotate_ctx(ctx);
+
+ perf_counter_cpu_sched_in(cpuctx, cpu);
+ if (ctx)
+ perf_counter_task_sched_in(curr, cpu);
}
/*
local_irq_save(flags);
if (ctx->is_active)
update_context_time(ctx);
- counter->hw_ops->read(counter);
+ counter->pmu->read(counter);
update_counter_times(counter);
local_irq_restore(flags);
}
return atomic64_read(&counter->count);
}
+/*
+ * Initialize the perf_counter context in a task_struct:
+ */
+static void
+__perf_counter_init_context(struct perf_counter_context *ctx,
+ struct task_struct *task)
+{
+ memset(ctx, 0, sizeof(*ctx));
+ spin_lock_init(&ctx->lock);
+ mutex_init(&ctx->mutex);
+ INIT_LIST_HEAD(&ctx->counter_list);
+ INIT_LIST_HEAD(&ctx->event_list);
+ atomic_set(&ctx->refcount, 1);
+ ctx->task = task;
+}
+
static void put_context(struct perf_counter_context *ctx)
{
if (ctx->task)
{
struct perf_cpu_context *cpuctx;
struct perf_counter_context *ctx;
+ struct perf_counter_context *tctx;
struct task_struct *task;
/*
*/
if (cpu != -1) {
/* Must be root to operate on a CPU counter: */
- if (!capable(CAP_SYS_ADMIN))
+ if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
return ERR_PTR(-EACCES);
if (cpu < 0 || cpu > num_possible_cpus())
if (!task)
return ERR_PTR(-ESRCH);
- ctx = &task->perf_counter_ctx;
- ctx->task = task;
-
/* Reuse ptrace permission checks for now. */
if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
- put_context(ctx);
+ put_task_struct(task);
return ERR_PTR(-EACCES);
}
+ ctx = task->perf_counter_ctxp;
+ if (!ctx) {
+ ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
+ if (!ctx) {
+ put_task_struct(task);
+ return ERR_PTR(-ENOMEM);
+ }
+ __perf_counter_init_context(ctx, task);
+ /*
+ * Make sure other cpus see correct values for *ctx
+ * once task->perf_counter_ctxp is visible to them.
+ */
+ smp_wmb();
+ tctx = cmpxchg(&task->perf_counter_ctxp, NULL, ctx);
+ if (tctx) {
+ /*
+ * We raced with some other task; use
+ * the context they set.
+ */
+ kfree(ctx);
+ ctx = tctx;
+ }
+ }
+
return ctx;
}
struct perf_counter *counter;
counter = container_of(head, struct perf_counter, rcu_head);
+ put_ctx(counter->ctx);
kfree(counter);
}
{
perf_pending_sync(counter);
+ atomic_dec(&nr_counters);
+ if (counter->hw_event.mmap)
+ atomic_dec(&nr_mmap_tracking);
+ if (counter->hw_event.munmap)
+ atomic_dec(&nr_munmap_tracking);
+ if (counter->hw_event.comm)
+ atomic_dec(&nr_comm_tracking);
+
if (counter->destroy)
counter->destroy(counter);
file->private_data = NULL;
mutex_lock(&ctx->mutex);
- mutex_lock(&counter->mutex);
-
perf_counter_remove_from_context(counter);
-
- mutex_unlock(&counter->mutex);
mutex_unlock(&ctx->mutex);
+ mutex_lock(&counter->owner->perf_counter_mutex);
+ list_del_init(&counter->owner_entry);
+ mutex_unlock(&counter->owner->perf_counter_mutex);
+ put_task_struct(counter->owner);
+
free_counter(counter);
put_context(ctx);
if (counter->state == PERF_COUNTER_STATE_ERROR)
return 0;
- mutex_lock(&counter->mutex);
+ mutex_lock(&counter->child_mutex);
values[0] = perf_counter_read(counter);
n = 1;
if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
values[n++] = counter->total_time_running +
atomic64_read(&counter->child_total_time_running);
- mutex_unlock(&counter->mutex);
+ mutex_unlock(&counter->child_mutex);
if (count < n * sizeof(u64))
return -EINVAL;
{
struct perf_counter *counter = file->private_data;
struct perf_mmap_data *data;
- unsigned int events;
+ unsigned int events = POLL_HUP;
rcu_read_lock();
data = rcu_dereference(counter->data);
if (data)
- events = atomic_xchg(&data->wakeup, 0);
- else
- events = POLL_HUP;
+ events = atomic_xchg(&data->poll, 0);
rcu_read_unlock();
poll_wait(file, &counter->waitq, wait);
return events;
}
+static void perf_counter_reset(struct perf_counter *counter)
+{
+ (void)perf_counter_read(counter);
+ atomic64_set(&counter->count, 0);
+ perf_counter_update_userpage(counter);
+}
+
+static void perf_counter_for_each_sibling(struct perf_counter *counter,
+ void (*func)(struct perf_counter *))
+{
+ struct perf_counter_context *ctx = counter->ctx;
+ struct perf_counter *sibling;
+
+ mutex_lock(&ctx->mutex);
+ counter = counter->group_leader;
+
+ func(counter);
+ list_for_each_entry(sibling, &counter->sibling_list, list_entry)
+ func(sibling);
+ mutex_unlock(&ctx->mutex);
+}
+
+static void perf_counter_for_each_child(struct perf_counter *counter,
+ void (*func)(struct perf_counter *))
+{
+ struct perf_counter *child;
+
+ mutex_lock(&counter->child_mutex);
+ func(counter);
+ list_for_each_entry(child, &counter->child_list, child_list)
+ func(child);
+ mutex_unlock(&counter->child_mutex);
+}
+
+static void perf_counter_for_each(struct perf_counter *counter,
+ void (*func)(struct perf_counter *))
+{
+ struct perf_counter *child;
+
+ mutex_lock(&counter->child_mutex);
+ perf_counter_for_each_sibling(counter, func);
+ list_for_each_entry(child, &counter->child_list, child_list)
+ perf_counter_for_each_sibling(child, func);
+ mutex_unlock(&counter->child_mutex);
+}
+
static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
struct perf_counter *counter = file->private_data;
- int err = 0;
+ void (*func)(struct perf_counter *);
+ u32 flags = arg;
switch (cmd) {
case PERF_COUNTER_IOC_ENABLE:
- perf_counter_enable_family(counter);
+ func = perf_counter_enable;
break;
case PERF_COUNTER_IOC_DISABLE:
- perf_counter_disable_family(counter);
+ func = perf_counter_disable;
break;
- case PERF_COUNTER_IOC_REFRESH:
- perf_counter_refresh(counter, arg);
+ case PERF_COUNTER_IOC_RESET:
+ func = perf_counter_reset;
break;
+
+ case PERF_COUNTER_IOC_REFRESH:
+ return perf_counter_refresh(counter, arg);
default:
- err = -ENOTTY;
+ return -ENOTTY;
}
- return err;
+
+ if (flags & PERF_IOC_FLAG_GROUP)
+ perf_counter_for_each(counter, func);
+ else
+ perf_counter_for_each_child(counter, func);
+
+ return 0;
}
/*
}
data->nr_pages = nr_pages;
+ atomic_set(&data->lock, -1);
rcu_assign_pointer(counter->data, data);
if (atomic_dec_and_mutex_lock(&counter->mmap_count,
&counter->mmap_mutex)) {
- vma->vm_mm->locked_vm -= counter->data->nr_pages + 1;
+ struct user_struct *user = current_user();
+
+ atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
+ vma->vm_mm->locked_vm -= counter->data->nr_locked;
perf_mmap_data_free(counter);
mutex_unlock(&counter->mmap_mutex);
}
static int perf_mmap(struct file *file, struct vm_area_struct *vma)
{
struct perf_counter *counter = file->private_data;
+ struct user_struct *user = current_user();
unsigned long vma_size;
unsigned long nr_pages;
+ unsigned long user_locked, user_lock_limit;
unsigned long locked, lock_limit;
+ long user_extra, extra;
int ret = 0;
if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
goto unlock;
}
- locked = vma->vm_mm->locked_vm;
- locked += nr_pages + 1;
+ user_extra = nr_pages + 1;
+ user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
+ user_locked = atomic_long_read(&user->locked_vm) + user_extra;
+
+ extra = 0;
+ if (user_locked > user_lock_limit)
+ extra = user_locked - user_lock_limit;
lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
lock_limit >>= PAGE_SHIFT;
+ locked = vma->vm_mm->locked_vm + extra;
if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
ret = -EPERM;
goto unlock;
atomic_set(&counter->mmap_count, 1);
- vma->vm_mm->locked_vm += nr_pages + 1;
+ atomic_long_add(user_extra, &user->locked_vm);
+ vma->vm_mm->locked_vm += extra;
+ counter->data->nr_locked = extra;
unlock:
mutex_unlock(&counter->mmap_mutex);
void perf_counter_wakeup(struct perf_counter *counter)
{
- struct perf_mmap_data *data;
-
- rcu_read_lock();
- data = rcu_dereference(counter->data);
- if (data) {
- atomic_set(&data->wakeup, POLL_IN);
- /*
- * Ensure all data writes are issued before updating the
- * user-space data head information. The matching rmb()
- * will be in userspace after reading this value.
- */
- smp_wmb();
- data->user_page->data_head = atomic_read(&data->head);
- }
- rcu_read_unlock();
-
wake_up_all(&counter->waitq);
if (counter->pending_kill) {
* Output
*/
-struct perf_output_handle {
- struct perf_counter *counter;
- struct perf_mmap_data *data;
- unsigned int offset;
- unsigned int head;
- int wakeup;
- int nmi;
- int overflow;
-};
+struct perf_output_handle {
+ struct perf_counter *counter;
+ struct perf_mmap_data *data;
+ unsigned int offset;
+ unsigned int head;
+ int nmi;
+ int overflow;
+ int locked;
+ unsigned long flags;
+};
+
+static void perf_output_wakeup(struct perf_output_handle *handle)
+{
+ atomic_set(&handle->data->poll, POLL_IN);
+
+ if (handle->nmi) {
+ handle->counter->pending_wakeup = 1;
+ perf_pending_queue(&handle->counter->pending,
+ perf_pending_counter);
+ } else
+ perf_counter_wakeup(handle->counter);
+}
+
+/*
+ * Curious locking construct.
+ *
+ * We need to ensure a later event doesn't publish a head when a former
+ * event isn't done writing. However since we need to deal with NMIs we
+ * cannot fully serialize things.
+ *
+ * What we do is serialize between CPUs so we only have to deal with NMI
+ * nesting on a single CPU.
+ *
+ * We only publish the head (and generate a wakeup) when the outer-most
+ * event completes.
+ */
+static void perf_output_lock(struct perf_output_handle *handle)
+{
+ struct perf_mmap_data *data = handle->data;
+ int cpu;
+
+ handle->locked = 0;
+
+ local_irq_save(handle->flags);
+ cpu = smp_processor_id();
+
+ if (in_nmi() && atomic_read(&data->lock) == cpu)
+ return;
+
+ while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
+ cpu_relax();
+
+ handle->locked = 1;
+}
+
+static void perf_output_unlock(struct perf_output_handle *handle)
+{
+ struct perf_mmap_data *data = handle->data;
+ int head, cpu;
+
+ data->done_head = data->head;
+
+ if (!handle->locked)
+ goto out;
+
+again:
+ /*
+ * The xchg implies a full barrier that ensures all writes are done
+ * before we publish the new head, matched by a rmb() in userspace when
+ * reading this position.
+ */
+ while ((head = atomic_xchg(&data->done_head, 0)))
+ data->user_page->data_head = head;
+
+ /*
+ * NMI can happen here, which means we can miss a done_head update.
+ */
+
+ cpu = atomic_xchg(&data->lock, -1);
+ WARN_ON_ONCE(cpu != smp_processor_id());
+
+ /*
+ * Therefore we have to validate we did not indeed do so.
+ */
+ if (unlikely(atomic_read(&data->done_head))) {
+ /*
+ * Since we had it locked, we can lock it again.
+ */
+ while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
+ cpu_relax();
+
+ goto again;
+ }
-static inline void __perf_output_wakeup(struct perf_output_handle *handle)
-{
- if (handle->nmi) {
- handle->counter->pending_wakeup = 1;
- perf_pending_queue(&handle->counter->pending,
- perf_pending_counter);
- } else
- perf_counter_wakeup(handle->counter);
+ if (atomic_xchg(&data->wakeup, 0))
+ perf_output_wakeup(handle);
+out:
+ local_irq_restore(handle->flags);
}
static int perf_output_begin(struct perf_output_handle *handle,
struct perf_mmap_data *data;
unsigned int offset, head;
+ /*
+ * For inherited counters we send all the output towards the parent.
+ */
+ if (counter->parent)
+ counter = counter->parent;
+
rcu_read_lock();
data = rcu_dereference(counter->data);
if (!data)
goto out;
+ handle->data = data;
handle->counter = counter;
handle->nmi = nmi;
handle->overflow = overflow;
if (!data->nr_pages)
goto fail;
+ perf_output_lock(handle);
+
do {
offset = head = atomic_read(&data->head);
head += size;
} while (atomic_cmpxchg(&data->head, offset, head) != offset);
- handle->data = data;
handle->offset = offset;
handle->head = head;
- handle->wakeup = (offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT);
+
+ if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
+ atomic_set(&data->wakeup, 1);
return 0;
fail:
- __perf_output_wakeup(handle);
+ perf_output_wakeup(handle);
out:
rcu_read_unlock();
handle->offset = offset;
- WARN_ON_ONCE(handle->offset > handle->head);
+ /*
+ * Check we didn't copy past our reservation window, taking the
+ * possible unsigned int wrap into account.
+ */
+ WARN_ON_ONCE(((int)(handle->head - handle->offset)) < 0);
}
#define perf_output_put(handle, x) \
static void perf_output_end(struct perf_output_handle *handle)
{
- int wakeup_events = handle->counter->hw_event.wakeup_events;
+ struct perf_counter *counter = handle->counter;
+ struct perf_mmap_data *data = handle->data;
+
+ int wakeup_events = counter->hw_event.wakeup_events;
if (handle->overflow && wakeup_events) {
- int events = atomic_inc_return(&handle->data->events);
+ int events = atomic_inc_return(&data->events);
if (events >= wakeup_events) {
- atomic_sub(wakeup_events, &handle->data->events);
- __perf_output_wakeup(handle);
+ atomic_sub(wakeup_events, &data->events);
+ atomic_set(&data->wakeup, 1);
}
- } else if (handle->wakeup)
- __perf_output_wakeup(handle);
+ }
+
+ perf_output_unlock(handle);
rcu_read_unlock();
}
static void perf_counter_output(struct perf_counter *counter,
- int nmi, struct pt_regs *regs)
+ int nmi, struct pt_regs *regs, u64 addr)
{
int ret;
u64 record_type = counter->hw_event.record_type;
struct perf_callchain_entry *callchain = NULL;
int callchain_size = 0;
u64 time;
+ struct {
+ u32 cpu, reserved;
+ } cpu_entry;
header.type = 0;
header.size = sizeof(header);
header.misc = PERF_EVENT_MISC_OVERFLOW;
- header.misc |= user_mode(regs) ?
- PERF_EVENT_MISC_USER : PERF_EVENT_MISC_KERNEL;
+ header.misc |= perf_misc_flags(regs);
if (record_type & PERF_RECORD_IP) {
- ip = instruction_pointer(regs);
+ ip = perf_instruction_pointer(regs);
header.type |= PERF_RECORD_IP;
header.size += sizeof(ip);
}
header.size += sizeof(tid_entry);
}
+ if (record_type & PERF_RECORD_TIME) {
+ /*
+ * Maybe do better on x86 and provide cpu_clock_nmi()
+ */
+ time = sched_clock();
+
+ header.type |= PERF_RECORD_TIME;
+ header.size += sizeof(u64);
+ }
+
+ if (record_type & PERF_RECORD_ADDR) {
+ header.type |= PERF_RECORD_ADDR;
+ header.size += sizeof(u64);
+ }
+
+ if (record_type & PERF_RECORD_CONFIG) {
+ header.type |= PERF_RECORD_CONFIG;
+ header.size += sizeof(u64);
+ }
+
+ if (record_type & PERF_RECORD_CPU) {
+ header.type |= PERF_RECORD_CPU;
+ header.size += sizeof(cpu_entry);
+
+ cpu_entry.cpu = raw_smp_processor_id();
+ }
+
if (record_type & PERF_RECORD_GROUP) {
header.type |= PERF_RECORD_GROUP;
header.size += sizeof(u64) +
}
}
- if (record_type & PERF_RECORD_TIME) {
- /*
- * Maybe do better on x86 and provide cpu_clock_nmi()
- */
- time = sched_clock();
-
- header.type |= PERF_RECORD_TIME;
- header.size += sizeof(u64);
- }
-
ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
if (ret)
return;
if (record_type & PERF_RECORD_TID)
perf_output_put(&handle, tid_entry);
+ if (record_type & PERF_RECORD_TIME)
+ perf_output_put(&handle, time);
+
+ if (record_type & PERF_RECORD_ADDR)
+ perf_output_put(&handle, addr);
+
+ if (record_type & PERF_RECORD_CONFIG)
+ perf_output_put(&handle, counter->hw_event.config);
+
+ if (record_type & PERF_RECORD_CPU)
+ perf_output_put(&handle, cpu_entry);
+
+ /*
+ * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
+ */
if (record_type & PERF_RECORD_GROUP) {
struct perf_counter *leader, *sub;
u64 nr = counter->nr_siblings;
leader = counter->group_leader;
list_for_each_entry(sub, &leader->sibling_list, list_entry) {
if (sub != counter)
- sub->hw_ops->read(sub);
+ sub->pmu->read(sub);
group_entry.event = sub->hw_event.config;
group_entry.counter = atomic64_read(&sub->count);
if (callchain)
perf_output_copy(&handle, callchain, callchain_size);
- if (record_type & PERF_RECORD_TIME)
- perf_output_put(&handle, time);
+ perf_output_end(&handle);
+}
+
+/*
+ * comm tracking
+ */
+
+struct perf_comm_event {
+ struct task_struct *task;
+ char *comm;
+ int comm_size;
+
+ struct {
+ struct perf_event_header header;
+
+ u32 pid;
+ u32 tid;
+ } event;
+};
+
+static void perf_counter_comm_output(struct perf_counter *counter,
+ struct perf_comm_event *comm_event)
+{
+ struct perf_output_handle handle;
+ int size = comm_event->event.header.size;
+ int ret = perf_output_begin(&handle, counter, size, 0, 0);
+
+ if (ret)
+ return;
+ perf_output_put(&handle, comm_event->event);
+ perf_output_copy(&handle, comm_event->comm,
+ comm_event->comm_size);
perf_output_end(&handle);
}
+static int perf_counter_comm_match(struct perf_counter *counter,
+ struct perf_comm_event *comm_event)
+{
+ if (counter->hw_event.comm &&
+ comm_event->event.header.type == PERF_EVENT_COMM)
+ return 1;
+
+ return 0;
+}
+
+static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
+ struct perf_comm_event *comm_event)
+{
+ struct perf_counter *counter;
+
+ if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
+ return;
+
+ rcu_read_lock();
+ list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
+ if (perf_counter_comm_match(counter, comm_event))
+ perf_counter_comm_output(counter, comm_event);
+ }
+ rcu_read_unlock();
+}
+
+static void perf_counter_comm_event(struct perf_comm_event *comm_event)
+{
+ struct perf_cpu_context *cpuctx;
+ unsigned int size;
+ char *comm = comm_event->task->comm;
+
+ size = ALIGN(strlen(comm)+1, sizeof(u64));
+
+ comm_event->comm = comm;
+ comm_event->comm_size = size;
+
+ comm_event->event.header.size = sizeof(comm_event->event) + size;
+
+ cpuctx = &get_cpu_var(perf_cpu_context);
+ perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
+ put_cpu_var(perf_cpu_context);
+
+ perf_counter_comm_ctx(current->perf_counter_ctxp, comm_event);
+}
+
+void perf_counter_comm(struct task_struct *task)
+{
+ struct perf_comm_event comm_event;
+
+ if (!atomic_read(&nr_comm_tracking))
+ return;
+ if (!current->perf_counter_ctxp)
+ return;
+
+ comm_event = (struct perf_comm_event){
+ .task = task,
+ .event = {
+ .header = { .type = PERF_EVENT_COMM, },
+ .pid = task->group_leader->pid,
+ .tid = task->pid,
+ },
+ };
+
+ perf_counter_comm_event(&comm_event);
+}
+
/*
* mmap tracking
*/
name = strncpy(tmp, "//enomem", sizeof(tmp));
goto got_name;
}
- name = dentry_path(file->f_dentry, buf, PATH_MAX);
+ name = d_path(&file->f_path, buf, PATH_MAX);
if (IS_ERR(name)) {
name = strncpy(tmp, "//toolong", sizeof(tmp));
goto got_name;
}
got_name:
- size = ALIGN(strlen(name), sizeof(u64));
+ size = ALIGN(strlen(name)+1, sizeof(u64));
mmap_event->file_name = name;
mmap_event->file_size = size;
perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
put_cpu_var(perf_cpu_context);
- perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
+ perf_counter_mmap_ctx(current->perf_counter_ctxp, mmap_event);
kfree(buf);
}
void perf_counter_mmap(unsigned long addr, unsigned long len,
unsigned long pgoff, struct file *file)
{
- struct perf_mmap_event mmap_event = {
+ struct perf_mmap_event mmap_event;
+
+ if (!atomic_read(&nr_mmap_tracking))
+ return;
+ if (!current->perf_counter_ctxp)
+ return;
+
+ mmap_event = (struct perf_mmap_event){
.file = file,
.event = {
.header = { .type = PERF_EVENT_MMAP, },
void perf_counter_munmap(unsigned long addr, unsigned long len,
unsigned long pgoff, struct file *file)
{
- struct perf_mmap_event mmap_event = {
+ struct perf_mmap_event mmap_event;
+
+ if (!atomic_read(&nr_munmap_tracking))
+ return;
+
+ mmap_event = (struct perf_mmap_event){
.file = file,
.event = {
.header = { .type = PERF_EVENT_MUNMAP, },
perf_counter_mmap_event(&mmap_event);
}
+/*
+ * Log irq_period changes so that analyzing tools can re-normalize the
+ * event flow.
+ */
+
+static void perf_log_period(struct perf_counter *counter, u64 period)
+{
+ struct perf_output_handle handle;
+ int ret;
+
+ struct {
+ struct perf_event_header header;
+ u64 time;
+ u64 period;
+ } freq_event = {
+ .header = {
+ .type = PERF_EVENT_PERIOD,
+ .misc = 0,
+ .size = sizeof(freq_event),
+ },
+ .time = sched_clock(),
+ .period = period,
+ };
+
+ if (counter->hw.irq_period == period)
+ return;
+
+ ret = perf_output_begin(&handle, counter, sizeof(freq_event), 0, 0);
+ if (ret)
+ return;
+
+ perf_output_put(&handle, freq_event);
+ perf_output_end(&handle);
+}
+
/*
* Generic counter overflow handling.
*/
int perf_counter_overflow(struct perf_counter *counter,
- int nmi, struct pt_regs *regs)
+ int nmi, struct pt_regs *regs, u64 addr)
{
int events = atomic_read(&counter->event_limit);
int ret = 0;
+ counter->hw.interrupts++;
+
+ /*
+ * XXX event_limit might not quite work as expected on inherited
+ * counters
+ */
+
counter->pending_kill = POLL_IN;
if (events && atomic_dec_and_test(&counter->event_limit)) {
ret = 1;
perf_counter_disable(counter);
}
- perf_counter_output(counter, nmi, regs);
+ perf_counter_output(counter, nmi, regs, addr);
return ret;
}
enum hrtimer_restart ret = HRTIMER_RESTART;
struct perf_counter *counter;
struct pt_regs *regs;
+ u64 period;
counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
- counter->hw_ops->read(counter);
+ counter->pmu->read(counter);
regs = get_irq_regs();
/*
regs = task_pt_regs(current);
if (regs) {
- if (perf_counter_overflow(counter, 0, regs))
+ if (perf_counter_overflow(counter, 0, regs, 0))
ret = HRTIMER_NORESTART;
}
- hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
+ period = max_t(u64, 10000, counter->hw.irq_period);
+ hrtimer_forward_now(hrtimer, ns_to_ktime(period));
return ret;
}
static void perf_swcounter_overflow(struct perf_counter *counter,
- int nmi, struct pt_regs *regs)
+ int nmi, struct pt_regs *regs, u64 addr)
{
perf_swcounter_update(counter);
perf_swcounter_set_period(counter);
- if (perf_counter_overflow(counter, nmi, regs))
+ if (perf_counter_overflow(counter, nmi, regs, addr))
/* soft-disable the counter */
;
}
static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
- int nmi, struct pt_regs *regs)
+ int nmi, struct pt_regs *regs, u64 addr)
{
int neg = atomic64_add_negative(nr, &counter->hw.count);
if (counter->hw.irq_period && !neg)
- perf_swcounter_overflow(counter, nmi, regs);
+ perf_swcounter_overflow(counter, nmi, regs, addr);
}
static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
enum perf_event_types type, u32 event,
- u64 nr, int nmi, struct pt_regs *regs)
+ u64 nr, int nmi, struct pt_regs *regs,
+ u64 addr)
{
struct perf_counter *counter;
rcu_read_lock();
list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
if (perf_swcounter_match(counter, type, event, regs))
- perf_swcounter_add(counter, nr, nmi, regs);
+ perf_swcounter_add(counter, nr, nmi, regs, addr);
}
rcu_read_unlock();
}
}
static void __perf_swcounter_event(enum perf_event_types type, u32 event,
- u64 nr, int nmi, struct pt_regs *regs)
+ u64 nr, int nmi, struct pt_regs *regs,
+ u64 addr)
{
struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
int *recursion = perf_swcounter_recursion_context(cpuctx);
(*recursion)++;
barrier();
- perf_swcounter_ctx_event(&cpuctx->ctx, type, event, nr, nmi, regs);
+ perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
+ nr, nmi, regs, addr);
if (cpuctx->task_ctx) {
perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
- nr, nmi, regs);
+ nr, nmi, regs, addr);
}
barrier();
put_cpu_var(perf_cpu_context);
}
-void perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)
+void
+perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
{
- __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs);
+ __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
}
static void perf_swcounter_read(struct perf_counter *counter)
perf_swcounter_update(counter);
}
-static const struct hw_perf_counter_ops perf_ops_generic = {
+static const struct pmu perf_ops_generic = {
.enable = perf_swcounter_enable,
.disable = perf_swcounter_disable,
.read = perf_swcounter_read,
hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
hwc->hrtimer.function = perf_swcounter_hrtimer;
if (hwc->irq_period) {
+ u64 period = max_t(u64, 10000, hwc->irq_period);
__hrtimer_start_range_ns(&hwc->hrtimer,
- ns_to_ktime(hwc->irq_period), 0,
+ ns_to_ktime(period), 0,
HRTIMER_MODE_REL, 0);
}
static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
{
- hrtimer_cancel(&counter->hw.hrtimer);
+ if (counter->hw.irq_period)
+ hrtimer_cancel(&counter->hw.hrtimer);
cpu_clock_perf_counter_update(counter);
}
cpu_clock_perf_counter_update(counter);
}
-static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
+static const struct pmu perf_ops_cpu_clock = {
.enable = cpu_clock_perf_counter_enable,
.disable = cpu_clock_perf_counter_disable,
.read = cpu_clock_perf_counter_read,
hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
hwc->hrtimer.function = perf_swcounter_hrtimer;
if (hwc->irq_period) {
+ u64 period = max_t(u64, 10000, hwc->irq_period);
__hrtimer_start_range_ns(&hwc->hrtimer,
- ns_to_ktime(hwc->irq_period), 0,
+ ns_to_ktime(period), 0,
HRTIMER_MODE_REL, 0);
}
static void task_clock_perf_counter_disable(struct perf_counter *counter)
{
- hrtimer_cancel(&counter->hw.hrtimer);
+ if (counter->hw.irq_period)
+ hrtimer_cancel(&counter->hw.hrtimer);
task_clock_perf_counter_update(counter, counter->ctx->time);
}
task_clock_perf_counter_update(counter, time);
}
-static const struct hw_perf_counter_ops perf_ops_task_clock = {
+static const struct pmu perf_ops_task_clock = {
.enable = task_clock_perf_counter_enable,
.disable = task_clock_perf_counter_disable,
.read = task_clock_perf_counter_read,
cpu_migrations_perf_counter_update(counter);
}
-static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
+static const struct pmu perf_ops_cpu_migrations = {
.enable = cpu_migrations_perf_counter_enable,
.disable = cpu_migrations_perf_counter_disable,
.read = cpu_migrations_perf_counter_read,
if (!regs)
regs = task_pt_regs(current);
- __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs);
+ __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
}
+EXPORT_SYMBOL_GPL(perf_tpcounter_event);
extern int ftrace_profile_enable(int);
extern void ftrace_profile_disable(int);
ftrace_profile_disable(perf_event_id(&counter->hw_event));
}
-static const struct hw_perf_counter_ops *
-tp_perf_counter_init(struct perf_counter *counter)
+static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
{
int event_id = perf_event_id(&counter->hw_event);
int ret;
return &perf_ops_generic;
}
#else
-static const struct hw_perf_counter_ops *
-tp_perf_counter_init(struct perf_counter *counter)
+static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
{
return NULL;
}
#endif
-static const struct hw_perf_counter_ops *
-sw_perf_counter_init(struct perf_counter *counter)
+static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
{
- struct perf_counter_hw_event *hw_event = &counter->hw_event;
- const struct hw_perf_counter_ops *hw_ops = NULL;
- struct hw_perf_counter *hwc = &counter->hw;
+ const struct pmu *pmu = NULL;
/*
* Software counters (currently) can't in general distinguish
*/
switch (perf_event_id(&counter->hw_event)) {
case PERF_COUNT_CPU_CLOCK:
- hw_ops = &perf_ops_cpu_clock;
+ pmu = &perf_ops_cpu_clock;
- if (hw_event->irq_period && hw_event->irq_period < 10000)
- hw_event->irq_period = 10000;
break;
case PERF_COUNT_TASK_CLOCK:
/*
* use the cpu_clock counter instead.
*/
if (counter->ctx->task)
- hw_ops = &perf_ops_task_clock;
+ pmu = &perf_ops_task_clock;
else
- hw_ops = &perf_ops_cpu_clock;
+ pmu = &perf_ops_cpu_clock;
- if (hw_event->irq_period && hw_event->irq_period < 10000)
- hw_event->irq_period = 10000;
break;
case PERF_COUNT_PAGE_FAULTS:
case PERF_COUNT_PAGE_FAULTS_MIN:
case PERF_COUNT_PAGE_FAULTS_MAJ:
case PERF_COUNT_CONTEXT_SWITCHES:
- hw_ops = &perf_ops_generic;
+ pmu = &perf_ops_generic;
break;
case PERF_COUNT_CPU_MIGRATIONS:
if (!counter->hw_event.exclude_kernel)
- hw_ops = &perf_ops_cpu_migrations;
+ pmu = &perf_ops_cpu_migrations;
break;
}
- if (hw_ops)
- hwc->irq_period = hw_event->irq_period;
-
- return hw_ops;
+ return pmu;
}
/*
struct perf_counter *group_leader,
gfp_t gfpflags)
{
- const struct hw_perf_counter_ops *hw_ops;
+ const struct pmu *pmu;
struct perf_counter *counter;
+ struct hw_perf_counter *hwc;
long err;
counter = kzalloc(sizeof(*counter), gfpflags);
if (!group_leader)
group_leader = counter;
- mutex_init(&counter->mutex);
+ mutex_init(&counter->child_mutex);
+ INIT_LIST_HEAD(&counter->child_list);
+
INIT_LIST_HEAD(&counter->list_entry);
INIT_LIST_HEAD(&counter->event_entry);
INIT_LIST_HEAD(&counter->sibling_list);
mutex_init(&counter->mmap_mutex);
- INIT_LIST_HEAD(&counter->child_list);
-
counter->cpu = cpu;
counter->hw_event = *hw_event;
counter->group_leader = group_leader;
- counter->hw_ops = NULL;
+ counter->pmu = NULL;
counter->ctx = ctx;
+ get_ctx(ctx);
counter->state = PERF_COUNTER_STATE_INACTIVE;
if (hw_event->disabled)
counter->state = PERF_COUNTER_STATE_OFF;
- hw_ops = NULL;
+ pmu = NULL;
+
+ hwc = &counter->hw;
+ if (hw_event->freq && hw_event->irq_freq)
+ hwc->irq_period = div64_u64(TICK_NSEC, hw_event->irq_freq);
+ else
+ hwc->irq_period = hw_event->irq_period;
+
+ /*
+ * we currently do not support PERF_RECORD_GROUP on inherited counters
+ */
+ if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
+ goto done;
if (perf_event_raw(hw_event)) {
- hw_ops = hw_perf_counter_init(counter);
+ pmu = hw_perf_counter_init(counter);
goto done;
}
switch (perf_event_type(hw_event)) {
case PERF_TYPE_HARDWARE:
- hw_ops = hw_perf_counter_init(counter);
+ pmu = hw_perf_counter_init(counter);
break;
case PERF_TYPE_SOFTWARE:
- hw_ops = sw_perf_counter_init(counter);
+ pmu = sw_perf_counter_init(counter);
break;
case PERF_TYPE_TRACEPOINT:
- hw_ops = tp_perf_counter_init(counter);
+ pmu = tp_perf_counter_init(counter);
break;
}
done:
err = 0;
- if (!hw_ops)
+ if (!pmu)
err = -EINVAL;
- else if (IS_ERR(hw_ops))
- err = PTR_ERR(hw_ops);
+ else if (IS_ERR(pmu))
+ err = PTR_ERR(pmu);
if (err) {
kfree(counter);
return ERR_PTR(err);
}
- counter->hw_ops = hw_ops;
+ counter->pmu = pmu;
+
+ atomic_inc(&nr_counters);
+ if (counter->hw_event.mmap)
+ atomic_inc(&nr_mmap_tracking);
+ if (counter->hw_event.munmap)
+ atomic_inc(&nr_munmap_tracking);
+ if (counter->hw_event.comm)
+ atomic_inc(&nr_comm_tracking);
return counter;
}
perf_install_in_context(ctx, counter, cpu);
mutex_unlock(&ctx->mutex);
+ counter->owner = current;
+ get_task_struct(current);
+ mutex_lock(¤t->perf_counter_mutex);
+ list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
+ mutex_unlock(¤t->perf_counter_mutex);
+
fput_light(counter_file, fput_needed2);
out_fput:
goto out_fput;
}
-/*
- * Initialize the perf_counter context in a task_struct:
- */
-static void
-__perf_counter_init_context(struct perf_counter_context *ctx,
- struct task_struct *task)
-{
- memset(ctx, 0, sizeof(*ctx));
- spin_lock_init(&ctx->lock);
- mutex_init(&ctx->mutex);
- INIT_LIST_HEAD(&ctx->counter_list);
- INIT_LIST_HEAD(&ctx->event_list);
- ctx->task = task;
-}
-
/*
* inherit a counter from parent task to child task:
*/
if (IS_ERR(child_counter))
return child_counter;
+ /*
+ * Make the child state follow the state of the parent counter,
+ * not its hw_event.disabled bit. We hold the parent's mutex,
+ * so we won't race with perf_counter_{en,dis}able_family.
+ */
+ if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
+ child_counter->state = PERF_COUNTER_STATE_INACTIVE;
+ else
+ child_counter->state = PERF_COUNTER_STATE_OFF;
+
/*
* Link it up in the child's context:
*/
- child_counter->task = child;
add_counter_to_ctx(child_counter, child_ctx);
child_counter->parent = parent_counter;
/*
* Link this into the parent counter's child list
*/
- mutex_lock(&parent_counter->mutex);
+ mutex_lock(&parent_counter->child_mutex);
list_add_tail(&child_counter->child_list, &parent_counter->child_list);
-
- /*
- * Make the child state follow the state of the parent counter,
- * not its hw_event.disabled bit. We hold the parent's mutex,
- * so we won't race with perf_counter_{en,dis}able_family.
- */
- if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
- child_counter->state = PERF_COUNTER_STATE_INACTIVE;
- else
- child_counter->state = PERF_COUNTER_STATE_OFF;
-
- mutex_unlock(&parent_counter->mutex);
+ mutex_unlock(&parent_counter->child_mutex);
return child_counter;
}
static void sync_child_counter(struct perf_counter *child_counter,
struct perf_counter *parent_counter)
{
- u64 parent_val, child_val;
+ u64 child_val;
- parent_val = atomic64_read(&parent_counter->count);
child_val = atomic64_read(&child_counter->count);
/*
/*
* Remove this counter from the parent's list
*/
- mutex_lock(&parent_counter->mutex);
+ mutex_lock(&parent_counter->child_mutex);
list_del_init(&child_counter->child_list);
- mutex_unlock(&parent_counter->mutex);
+ mutex_unlock(&parent_counter->child_mutex);
/*
* Release the parent counter, if this was the last
struct perf_counter_context *child_ctx)
{
struct perf_counter *parent_counter;
- struct perf_counter *sub, *tmp;
-
- /*
- * If we do not self-reap then we have to wait for the
- * child task to unschedule (it will happen for sure),
- * so that its counter is at its final count. (This
- * condition triggers rarely - child tasks usually get
- * off their CPU before the parent has a chance to
- * get this far into the reaping action)
- */
- if (child != current) {
- wait_task_inactive(child, 0);
- list_del_init(&child_counter->list_entry);
- update_counter_times(child_counter);
- } else {
- struct perf_cpu_context *cpuctx;
- unsigned long flags;
- u64 perf_flags;
-
- /*
- * Disable and unlink this counter.
- *
- * Be careful about zapping the list - IRQ/NMI context
- * could still be processing it:
- */
- local_irq_save(flags);
- perf_flags = hw_perf_save_disable();
-
- cpuctx = &__get_cpu_var(perf_cpu_context);
-
- group_sched_out(child_counter, cpuctx, child_ctx);
- update_counter_times(child_counter);
-
- list_del_init(&child_counter->list_entry);
-
- child_ctx->nr_counters--;
- hw_perf_restore(perf_flags);
- local_irq_restore(flags);
- }
+ update_counter_times(child_counter);
+ perf_counter_remove_from_context(child_counter);
parent_counter = child_counter->parent;
/*
*/
if (parent_counter) {
sync_child_counter(child_counter, parent_counter);
- list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
- list_entry) {
- if (sub->parent) {
- sync_child_counter(sub, sub->parent);
- free_counter(sub);
- }
- }
free_counter(child_counter);
}
}
*
* Note: we may be running in child context, but the PID is not hashed
* anymore so new counters will not be added.
+ * (XXX not sure that is true when we get called from flush_old_exec.
+ * -- paulus)
*/
void perf_counter_exit_task(struct task_struct *child)
{
struct perf_counter *child_counter, *tmp;
struct perf_counter_context *child_ctx;
+ unsigned long flags;
- child_ctx = &child->perf_counter_ctx;
+ WARN_ON_ONCE(child != current);
- if (likely(!child_ctx->nr_counters))
+ child_ctx = child->perf_counter_ctxp;
+
+ if (likely(!child_ctx))
return;
+ local_irq_save(flags);
+ __perf_counter_task_sched_out(child_ctx);
+ child->perf_counter_ctxp = NULL;
+ local_irq_restore(flags);
+
+ mutex_lock(&child_ctx->mutex);
+
+again:
list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
list_entry)
__perf_counter_exit_task(child, child_counter, child_ctx);
+
+ /*
+ * If the last counter was a group counter, it will have appended all
+ * its siblings to the list, but we obtained 'tmp' before that which
+ * will still point to the list head terminating the iteration.
+ */
+ if (!list_empty(&child_ctx->counter_list))
+ goto again;
+
+ mutex_unlock(&child_ctx->mutex);
+
+ put_ctx(child_ctx);
}
/*
struct perf_counter_context *child_ctx, *parent_ctx;
struct perf_counter *counter;
struct task_struct *parent = current;
+ int inherited_all = 1;
- child_ctx = &child->perf_counter_ctx;
- parent_ctx = &parent->perf_counter_ctx;
+ child->perf_counter_ctxp = NULL;
- __perf_counter_init_context(child_ctx, child);
+ mutex_init(&child->perf_counter_mutex);
+ INIT_LIST_HEAD(&child->perf_counter_list);
/*
* This is executed from the parent task context, so inherit
- * counters that have been marked for cloning:
+ * counters that have been marked for cloning.
+ * First allocate and initialize a context for the child.
*/
- if (likely(!parent_ctx->nr_counters))
+ child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
+ if (!child_ctx)
+ return;
+
+ parent_ctx = parent->perf_counter_ctxp;
+ if (likely(!parent_ctx || !parent_ctx->nr_counters))
return;
+ __perf_counter_init_context(child_ctx, child);
+ child->perf_counter_ctxp = child_ctx;
+
/*
* Lock the parent list. No need to lock the child - not PID
* hashed yet and not running, so nobody can access it.
* We dont have to disable NMIs - we are only looking at
* the list, not manipulating it:
*/
- list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
- if (!counter->hw_event.inherit)
+ list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
+ if (counter != counter->group_leader)
+ continue;
+
+ if (!counter->hw_event.inherit) {
+ inherited_all = 0;
continue;
+ }
if (inherit_group(counter, parent,
- parent_ctx, child, child_ctx))
+ parent_ctx, child, child_ctx)) {
+ inherited_all = 0;
break;
+ }
+ }
+
+ if (inherited_all) {
+ /*
+ * Mark the child context as a clone of the parent
+ * context, or of whatever the parent is a clone of.
+ */
+ if (parent_ctx->parent_ctx) {
+ child_ctx->parent_ctx = parent_ctx->parent_ctx;
+ child_ctx->parent_gen = parent_ctx->parent_gen;
+ } else {
+ child_ctx->parent_ctx = parent_ctx;
+ child_ctx->parent_gen = parent_ctx->generation;
+ }
+ get_ctx(child_ctx->parent_ctx);
}
mutex_unlock(&parent_ctx->mutex);
cpuctx = &per_cpu(perf_cpu_context, cpu);
__perf_counter_init_context(&cpuctx->ctx, NULL);
- mutex_lock(&perf_resource_mutex);
+ spin_lock(&perf_resource_lock);
cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
- mutex_unlock(&perf_resource_mutex);
+ spin_unlock(&perf_resource_lock);
hw_perf_counter_setup(cpu);
}
.notifier_call = perf_cpu_notify,
};
-static int __init perf_counter_init(void)
+void __init perf_counter_init(void)
{
perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
(void *)(long)smp_processor_id());
register_cpu_notifier(&perf_cpu_nb);
-
- return 0;
}
-early_initcall(perf_counter_init);
static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
{
if (val > perf_max_counters)
return -EINVAL;
- mutex_lock(&perf_resource_mutex);
+ spin_lock(&perf_resource_lock);
perf_reserved_percpu = val;
for_each_online_cpu(cpu) {
cpuctx = &per_cpu(perf_cpu_context, cpu);
cpuctx->max_pertask = mpt;
spin_unlock_irq(&cpuctx->ctx.lock);
}
- mutex_unlock(&perf_resource_mutex);
+ spin_unlock(&perf_resource_lock);
return count;
}
if (val > 1)
return -EINVAL;
- mutex_lock(&perf_resource_mutex);
+ spin_lock(&perf_resource_lock);
perf_overcommit = val;
- mutex_unlock(&perf_resource_mutex);
+ spin_unlock(&perf_resource_lock);
return count;
}