#include <linux/file.h>
#include <linux/poll.h>
#include <linux/sysfs.h>
-#include <linux/ptrace.h>
+#include <linux/dcache.h>
#include <linux/percpu.h>
+#include <linux/ptrace.h>
#include <linux/vmstat.h>
#include <linux/hardirq.h>
#include <linux/rculist.h>
#include <linux/anon_inodes.h>
#include <linux/kernel_stat.h>
#include <linux/perf_counter.h>
-#include <linux/dcache.h>
#include <asm/irq_regs.h>
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;
+static atomic_t nr_mmap_counters __read_mostly;
+static atomic_t nr_munmap_counters __read_mostly;
+static atomic_t nr_comm_counters __read_mostly;
int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
-int sysctl_perf_counter_mlock __read_mostly = 128; /* 'free' kb per counter */
+int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
+int sysctl_perf_counter_limit __read_mostly = 100000; /* max NMIs per second */
+
+static atomic64_t perf_counter_id;
/*
* Lock for (sysadmin-configurable) counter reservations:
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,
+
+int __weak
+hw_perf_group_sched_in(struct perf_counter *group_leader,
struct perf_cpu_context *cpuctx,
struct perf_counter_context *ctx, int cpu)
{
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 free_ctx(struct rcu_head *head)
+{
+ struct perf_counter_context *ctx;
+
+ ctx = container_of(head, struct perf_counter_context, rcu_head);
+ kfree(ctx);
+}
+
+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);
+ if (ctx->task)
+ put_task_struct(ctx->task);
+ call_rcu(&ctx->rcu_head, free_ctx);
+ }
+}
+
+/*
+ * Get the perf_counter_context for a task and lock it.
+ * This has to cope with with the fact that until it is locked,
+ * the context could get moved to another task.
+ */
+static struct perf_counter_context *
+perf_lock_task_context(struct task_struct *task, unsigned long *flags)
+{
+ struct perf_counter_context *ctx;
+
+ rcu_read_lock();
+ retry:
+ ctx = rcu_dereference(task->perf_counter_ctxp);
+ if (ctx) {
+ /*
+ * If this context is a clone of another, it might
+ * get swapped for another underneath us by
+ * perf_counter_task_sched_out, though the
+ * rcu_read_lock() protects us from any context
+ * getting freed. Lock the context and check if it
+ * got swapped before we could get the lock, and retry
+ * if so. If we locked the right context, then it
+ * can't get swapped on us any more.
+ */
+ spin_lock_irqsave(&ctx->lock, *flags);
+ if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
+ spin_unlock_irqrestore(&ctx->lock, *flags);
+ goto retry;
+ }
+ }
+ rcu_read_unlock();
+ return ctx;
+}
+
+/*
+ * Get the context for a task and increment its pin_count so it
+ * can't get swapped to another task. This also increments its
+ * reference count so that the context can't get freed.
+ */
+static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
+{
+ struct perf_counter_context *ctx;
+ unsigned long flags;
+
+ ctx = perf_lock_task_context(task, &flags);
+ if (ctx) {
+ ++ctx->pin_count;
+ get_ctx(ctx);
+ spin_unlock_irqrestore(&ctx->lock, flags);
+ }
+ return ctx;
+}
+
+static void perf_unpin_context(struct perf_counter_context *ctx)
+{
+ unsigned long flags;
+
+ spin_lock_irqsave(&ctx->lock, flags);
+ --ctx->pin_count;
+ spin_unlock_irqrestore(&ctx->lock, flags);
+ put_ctx(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)
{
}
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);
if (!is_software_counter(counter))
cpuctx->active_oncpu--;
ctx->nr_active--;
- if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
+ if (counter->attr.exclusive || !cpuctx->active_oncpu)
cpuctx->exclusive = 0;
}
list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
counter_sched_out(counter, cpuctx, ctx);
- if (group_counter->hw_event.exclusive)
+ if (group_counter->attr.exclusive)
cpuctx->exclusive = 0;
}
struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
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
if (ctx->task && cpuctx->task_ctx != ctx)
return;
- spin_lock_irqsave(&ctx->lock, flags);
-
- counter_sched_out(counter, cpuctx, ctx);
-
- counter->task = NULL;
- ctx->nr_counters--;
-
+ spin_lock(&ctx->lock);
/*
* 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);
}
- spin_unlock_irqrestore(&ctx->lock, flags);
+ perf_enable();
+ spin_unlock(&ctx->lock);
}
/*
* 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.
+ *
+ * If counter->ctx is a cloned context, callers must make sure that
+ * every task struct that counter->ctx->task could possibly point to
+ * remains valid. This is OK when called from perf_release since
+ * that only calls us on the top-level context, which can't be a clone.
+ * When called from perf_counter_exit_task, it's OK because the
+ * context has been detached from its task.
*/
static void perf_counter_remove_from_context(struct perf_counter *counter)
{
* succeed.
*/
if (!list_empty(&counter->list_entry)) {
- ctx->nr_counters--;
list_del_counter(counter, ctx);
- counter->task = NULL;
}
spin_unlock_irq(&ctx->lock);
}
struct perf_counter *counter = info;
struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
struct perf_counter_context *ctx = counter->ctx;
- unsigned long flags;
/*
* If this is a per-task counter, need to check whether this
if (ctx->task && cpuctx->task_ctx != ctx)
return;
- spin_lock_irqsave(&ctx->lock, flags);
+ spin_lock(&ctx->lock);
/*
* If the counter is on, turn it off.
counter->state = PERF_COUNTER_STATE_OFF;
}
- spin_unlock_irqrestore(&ctx->lock, flags);
+ spin_unlock(&ctx->lock);
}
/*
* Disable a counter.
+ *
+ * If counter->ctx is a cloned context, callers must make sure that
+ * every task struct that counter->ctx->task could possibly point to
+ * remains valid. This condition is satisifed when called through
+ * perf_counter_for_each_child or perf_counter_for_each because they
+ * hold the top-level counter's child_mutex, so any descendant that
+ * goes to exit will block in sync_child_counter.
+ * When called from perf_pending_counter it's OK because counter->ctx
+ * is the current context on this CPU and preemption is disabled,
+ * hence we can't get into perf_counter_task_sched_out for this context.
*/
static void perf_counter_disable(struct perf_counter *counter)
{
cpuctx->active_oncpu++;
ctx->nr_active++;
- if (counter->hw_event.exclusive)
+ if (counter->attr.exclusive)
cpuctx->exclusive = 1;
return 0;
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;
* If this group is exclusive and there are already
* counters on the CPU, it can't go on.
*/
- if (counter->hw_event.exclusive && cpuctx->active_oncpu)
+ if (counter->attr.exclusive && cpuctx->active_oncpu)
return 0;
/*
* Otherwise, try to add it if all previous groups were able
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;
counter->tstamp_stopped = 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_context *ctx = counter->ctx;
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);
+ spin_lock(&ctx->lock);
+ 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);
*/
if (leader != counter)
group_sched_out(leader, cpuctx, ctx);
- if (leader->hw_event.pinned) {
+ if (leader->attr.pinned) {
update_group_times(leader);
leader->state = PERF_COUNTER_STATE_ERROR;
}
cpuctx->max_pertask--;
unlock:
- hw_perf_restore(perf_flags);
+ perf_enable();
- spin_unlock_irqrestore(&ctx->lock, flags);
+ spin_unlock(&ctx->lock);
}
/*
return;
}
- counter->task = task;
retry:
task_oncpu_function_call(task, __perf_install_in_context,
counter);
struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
struct perf_counter_context *ctx = counter->ctx;
struct perf_counter *leader = counter->group_leader;
- unsigned long pmuflags;
- unsigned long flags;
int err;
/*
* 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);
+ spin_lock(&ctx->lock);
+ ctx->is_active = 1;
update_context_time(ctx);
- counter->prev_state = counter->state;
if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
goto unlock;
counter->state = PERF_COUNTER_STATE_INACTIVE;
if (!group_can_go_on(counter, cpuctx, 1)) {
err = -EEXIST;
} else {
- pmuflags = hw_perf_save_disable();
+ 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());
- hw_perf_restore(pmuflags);
+ perf_enable();
}
if (err) {
*/
if (leader != counter)
group_sched_out(leader, cpuctx, ctx);
- if (leader->hw_event.pinned) {
+ if (leader->attr.pinned) {
update_group_times(leader);
leader->state = PERF_COUNTER_STATE_ERROR;
}
}
unlock:
- spin_unlock_irqrestore(&ctx->lock, flags);
+ spin_unlock(&ctx->lock);
}
/*
* Enable a counter.
+ *
+ * If counter->ctx is a cloned context, callers must make sure that
+ * every task struct that counter->ctx->task could possibly point to
+ * remains valid. This condition is satisfied when called through
+ * perf_counter_for_each_child or perf_counter_for_each as described
+ * for perf_counter_disable.
*/
static void perf_counter_enable(struct perf_counter *counter)
{
/*
* not supported on inherited counters
*/
- if (counter->hw_event.inherit)
+ if (counter->attr.inherit)
return -EINVAL;
atomic_add(refresh, &counter->event_limit);
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
+ && !ctx1->pin_count && !ctx2->pin_count;
+}
+
/*
* 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 perf_counter_context *parent;
struct pt_regs *regs;
+ int do_switch = 1;
+
+ regs = task_pt_regs(task);
+ perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
- 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, 0);
- __perf_counter_sched_out(ctx, cpuctx);
+ rcu_read_lock();
+ parent = rcu_dereference(ctx->parent_ctx);
+ next_ctx = next->perf_counter_ctxp;
+ if (parent && next_ctx &&
+ rcu_dereference(next_ctx->parent_ctx) == parent) {
+ /*
+ * Looks like the two contexts are clones, so we might be
+ * able to optimize the context switch. We lock both
+ * contexts and check that they are clones under the
+ * lock (including re-checking that neither has been
+ * uncloned in the meantime). It doesn't matter which
+ * order we take the locks because no other cpu could
+ * be trying to lock both of these tasks.
+ */
+ spin_lock(&ctx->lock);
+ spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
+ if (context_equiv(ctx, next_ctx)) {
+ /*
+ * XXX do we need a memory barrier of sorts
+ * wrt to rcu_dereference() of perf_counter_ctxp
+ */
+ task->perf_counter_ctxp = next_ctx;
+ next->perf_counter_ctxp = ctx;
+ ctx->task = next;
+ next_ctx->task = task;
+ do_switch = 0;
+ }
+ spin_unlock(&next_ctx->lock);
+ spin_unlock(&ctx->lock);
+ }
+ rcu_read_unlock();
- cpuctx->task_ctx = NULL;
+ if (do_switch) {
+ __perf_counter_sched_out(ctx, cpuctx);
+ cpuctx->task_ctx = NULL;
+ }
}
+/*
+ * Called with IRQs disabled
+ */
static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
{
struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
+ if (!cpuctx->task_ctx)
+ return;
+
+ if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
+ return;
+
__perf_counter_sched_out(ctx, cpuctx);
cpuctx->task_ctx = NULL;
}
+/*
+ * Called with IRQs disabled
+ */
static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
{
__perf_counter_sched_out(&cpuctx->ctx, cpuctx);
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
*/
list_for_each_entry(counter, &ctx->counter_list, list_entry) {
if (counter->state <= PERF_COUNTER_STATE_OFF ||
- !counter->hw_event.pinned)
+ !counter->attr.pinned)
continue;
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,
* ignore pinned counters since we did them already.
*/
if (counter->state <= PERF_COUNTER_STATE_OFF ||
- counter->hw_event.pinned)
+ counter->attr.pinned)
continue;
/*
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)
-{
- struct task_struct *curr = current;
- struct perf_counter_context *ctx = &curr->perf_counter_ctx;
- struct perf_counter *counter;
- unsigned long flags;
- u64 perf_flags;
-
- if (likely(!ctx->nr_counters))
- return 0;
+#define MAX_INTERRUPTS (~0ULL)
- local_irq_save(flags);
+static void perf_log_throttle(struct perf_counter *counter, int enable);
+static void perf_log_period(struct perf_counter *counter, u64 period);
- __perf_counter_task_sched_out(ctx);
+static void perf_adjust_freq(struct perf_counter_context *ctx)
+{
+ struct perf_counter *counter;
+ u64 interrupts, sample_period;
+ u64 events, period;
+ s64 delta;
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);
+ if (counter->state != PERF_COUNTER_STATE_ACTIVE)
+ continue;
- spin_unlock_irqrestore(&ctx->lock, flags);
+ interrupts = counter->hw.interrupts;
+ counter->hw.interrupts = 0;
- return 0;
-}
+ if (interrupts == MAX_INTERRUPTS) {
+ perf_log_throttle(counter, 1);
+ counter->pmu->unthrottle(counter);
+ interrupts = 2*sysctl_perf_counter_limit/HZ;
+ }
-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 (!counter->attr.freq || !counter->attr.sample_freq)
+ continue;
- if (likely(!ctx->nr_counters))
- return 0;
+ events = HZ * interrupts * counter->hw.sample_period;
+ period = div64_u64(events, counter->attr.sample_freq);
- local_irq_save(flags);
- cpu = smp_processor_id();
+ delta = (s64)(1 + period - counter->hw.sample_period);
+ delta >>= 1;
- __perf_counter_task_sched_out(ctx);
+ sample_period = counter->hw.sample_period + delta;
- spin_lock(&ctx->lock);
+ if (!sample_period)
+ sample_period = 1;
- /*
- * Disable all the counters:
- */
- perf_flags = hw_perf_save_disable();
+ perf_log_period(counter, sample_period);
- list_for_each_entry(counter, &ctx->counter_list, list_entry) {
- if (counter->state > PERF_COUNTER_STATE_OFF)
- continue;
- counter->state = PERF_COUNTER_STATE_INACTIVE;
- counter->tstamp_enabled =
- ctx->time - counter->total_time_enabled;
- counter->hw_event.disabled = 0;
+ counter->hw.sample_period = sample_period;
}
- hw_perf_restore(perf_flags);
-
spin_unlock(&ctx->lock);
-
- perf_counter_task_sched_in(curr, cpu);
-
- local_irq_restore(flags);
-
- return 0;
}
/*
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);
}
return;
cpuctx = &per_cpu(perf_cpu_context, cpu);
- ctx = &curr->perf_counter_ctx;
+ ctx = curr->perf_counter_ctxp;
+
+ perf_adjust_freq(&cpuctx->ctx);
+ if (ctx)
+ perf_adjust_freq(ctx);
perf_counter_cpu_sched_out(cpuctx);
- __perf_counter_task_sched_out(ctx);
+ if (ctx)
+ __perf_counter_task_sched_out(ctx);
rotate_ctx(&cpuctx->ctx);
- rotate_ctx(ctx);
+ if (ctx)
+ rotate_ctx(ctx);
perf_counter_cpu_sched_in(cpuctx, cpu);
- perf_counter_task_sched_in(curr, cpu);
+ if (ctx)
+ perf_counter_task_sched_in(curr, cpu);
}
/*
return atomic64_read(&counter->count);
}
-static void put_context(struct perf_counter_context *ctx)
+/*
+ * Initialize the perf_counter context in a task_struct:
+ */
+static void
+__perf_counter_init_context(struct perf_counter_context *ctx,
+ struct task_struct *task)
{
- if (ctx->task)
- put_task_struct(ctx->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 struct perf_counter_context *find_get_context(pid_t pid, int cpu)
{
- struct perf_cpu_context *cpuctx;
+ struct perf_counter_context *parent_ctx;
struct perf_counter_context *ctx;
+ struct perf_cpu_context *cpuctx;
struct task_struct *task;
+ unsigned long flags;
+ int err;
/*
* If cpu is not a wildcard then this is a percpu counter:
cpuctx = &per_cpu(perf_cpu_context, cpu);
ctx = &cpuctx->ctx;
+ get_ctx(ctx);
return ctx;
}
if (!task)
return ERR_PTR(-ESRCH);
- ctx = &task->perf_counter_ctx;
- ctx->task = task;
+ /*
+ * Can't attach counters to a dying task.
+ */
+ err = -ESRCH;
+ if (task->flags & PF_EXITING)
+ goto errout;
/* Reuse ptrace permission checks for now. */
- if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
- put_context(ctx);
- return ERR_PTR(-EACCES);
+ err = -EACCES;
+ if (!ptrace_may_access(task, PTRACE_MODE_READ))
+ goto errout;
+
+ retry:
+ ctx = perf_lock_task_context(task, &flags);
+ if (ctx) {
+ parent_ctx = ctx->parent_ctx;
+ if (parent_ctx) {
+ put_ctx(parent_ctx);
+ ctx->parent_ctx = NULL; /* no longer a clone */
+ }
+ /*
+ * Get an extra reference before dropping the lock so that
+ * this context won't get freed if the task exits.
+ */
+ get_ctx(ctx);
+ spin_unlock_irqrestore(&ctx->lock, flags);
+ }
+
+ if (!ctx) {
+ ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
+ err = -ENOMEM;
+ if (!ctx)
+ goto errout;
+ __perf_counter_init_context(ctx, task);
+ get_ctx(ctx);
+ if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
+ /*
+ * We raced with some other task; use
+ * the context they set.
+ */
+ kfree(ctx);
+ goto retry;
+ }
+ get_task_struct(task);
}
+ put_task_struct(task);
return ctx;
+
+ errout:
+ put_task_struct(task);
+ return ERR_PTR(err);
}
static void free_counter_rcu(struct rcu_head *head)
struct perf_counter *counter;
counter = container_of(head, struct perf_counter, rcu_head);
+ if (counter->ns)
+ put_pid_ns(counter->ns);
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->attr.mmap)
+ atomic_dec(&nr_mmap_counters);
+ if (counter->attr.munmap)
+ atomic_dec(&nr_munmap_counters);
+ if (counter->attr.comm)
+ atomic_dec(&nr_comm_counters);
if (counter->destroy)
counter->destroy(counter);
+ put_ctx(counter->ctx);
call_rcu(&counter->rcu_head, free_counter_rcu);
}
file->private_data = NULL;
+ WARN_ON_ONCE(ctx->parent_ctx);
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);
return 0;
}
if (counter->state == PERF_COUNTER_STATE_ERROR)
return 0;
- mutex_lock(&counter->mutex);
+ WARN_ON_ONCE(counter->ctx->parent_ctx);
+ 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->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
values[n++] = counter->total_time_enabled +
atomic64_read(&counter->child_total_time_enabled);
- if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
+ if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
values[n++] = counter->total_time_running +
atomic64_read(&counter->child_total_time_running);
- mutex_unlock(&counter->mutex);
+ if (counter->attr.read_format & PERF_FORMAT_ID)
+ values[n++] = counter->id;
+ mutex_unlock(&counter->child_mutex);
if (count < n * sizeof(u64))
return -EINVAL;
struct perf_counter_context *ctx = counter->ctx;
struct perf_counter *sibling;
- spin_lock_irq(&ctx->lock);
+ WARN_ON_ONCE(ctx->parent_ctx);
+ mutex_lock(&ctx->mutex);
counter = counter->group_leader;
func(counter);
list_for_each_entry(sibling, &counter->sibling_list, list_entry)
func(sibling);
- spin_unlock_irq(&ctx->lock);
+ mutex_unlock(&ctx->mutex);
}
+/*
+ * Holding the top-level counter's child_mutex means that any
+ * descendant process that has inherited this counter will block
+ * in sync_child_counter if it goes to exit, thus satisfying the
+ * task existence requirements of perf_counter_enable/disable.
+ */
static void perf_counter_for_each_child(struct perf_counter *counter,
void (*func)(struct perf_counter *))
{
struct perf_counter *child;
- mutex_lock(&counter->mutex);
+ WARN_ON_ONCE(counter->ctx->parent_ctx);
+ mutex_lock(&counter->child_mutex);
func(counter);
list_for_each_entry(child, &counter->child_list, child_list)
func(child);
- mutex_unlock(&counter->mutex);
+ mutex_unlock(&counter->child_mutex);
}
static void perf_counter_for_each(struct perf_counter *counter,
{
struct perf_counter *child;
- mutex_lock(&counter->mutex);
+ WARN_ON_ONCE(counter->ctx->parent_ctx);
+ 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->mutex);
+ mutex_unlock(&counter->child_mutex);
}
-static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
{
- struct perf_counter *counter = file->private_data;
- void (*func)(struct perf_counter *);
- u32 flags = arg;
+ struct perf_counter_context *ctx = counter->ctx;
+ unsigned long size;
+ int ret = 0;
+ u64 value;
- switch (cmd) {
- case PERF_COUNTER_IOC_ENABLE:
- func = perf_counter_enable;
- break;
- case PERF_COUNTER_IOC_DISABLE:
- func = perf_counter_disable;
- break;
- case PERF_COUNTER_IOC_RESET:
- func = perf_counter_reset;
- break;
+ if (!counter->attr.sample_period)
+ return -EINVAL;
+
+ size = copy_from_user(&value, arg, sizeof(value));
+ if (size != sizeof(value))
+ return -EFAULT;
+
+ if (!value)
+ return -EINVAL;
+
+ spin_lock_irq(&ctx->lock);
+ if (counter->attr.freq) {
+ if (value > sysctl_perf_counter_limit) {
+ ret = -EINVAL;
+ goto unlock;
+ }
+
+ counter->attr.sample_freq = value;
+ } else {
+ counter->attr.sample_period = value;
+ counter->hw.sample_period = value;
+
+ perf_log_period(counter, value);
+ }
+unlock:
+ spin_unlock_irq(&ctx->lock);
+
+ return ret;
+}
+
+static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+{
+ struct perf_counter *counter = file->private_data;
+ void (*func)(struct perf_counter *);
+ u32 flags = arg;
+
+ switch (cmd) {
+ case PERF_COUNTER_IOC_ENABLE:
+ func = perf_counter_enable;
+ break;
+ case PERF_COUNTER_IOC_DISABLE:
+ func = perf_counter_disable;
+ break;
+ case PERF_COUNTER_IOC_RESET:
+ func = perf_counter_reset;
+ break;
+
+ case PERF_COUNTER_IOC_REFRESH:
+ return perf_counter_refresh(counter, arg);
+
+ case PERF_COUNTER_IOC_PERIOD:
+ return perf_counter_period(counter, (u64 __user *)arg);
- case PERF_COUNTER_IOC_REFRESH:
- return perf_counter_refresh(counter, arg);
default:
return -ENOTTY;
}
return 0;
}
+int perf_counter_task_enable(void)
+{
+ struct perf_counter *counter;
+
+ mutex_lock(¤t->perf_counter_mutex);
+ list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
+ perf_counter_for_each_child(counter, perf_counter_enable);
+ mutex_unlock(¤t->perf_counter_mutex);
+
+ return 0;
+}
+
+int perf_counter_task_disable(void)
+{
+ struct perf_counter *counter;
+
+ mutex_lock(¤t->perf_counter_mutex);
+ list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
+ perf_counter_for_each_child(counter, perf_counter_disable);
+ mutex_unlock(¤t->perf_counter_mutex);
+
+ return 0;
+}
+
/*
* 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
*/
void perf_counter_update_userpage(struct perf_counter *counter)
{
- struct perf_mmap_data *data;
struct perf_counter_mmap_page *userpg;
+ struct perf_mmap_data *data;
rcu_read_lock();
data = rcu_dereference(counter->data);
static void __perf_mmap_data_free(struct rcu_head *rcu_head)
{
- struct perf_mmap_data *data = container_of(rcu_head,
- struct perf_mmap_data, rcu_head);
+ struct perf_mmap_data *data;
int i;
+ data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
+
free_page((unsigned long)data->user_page);
for (i = 0; i < data->nr_pages; i++)
free_page((unsigned long)data->data_pages[i]);
{
struct perf_counter *counter = vma->vm_file->private_data;
- if (atomic_dec_and_mutex_lock(&counter->mmap_count,
- &counter->mmap_mutex)) {
+ WARN_ON_ONCE(counter->ctx->parent_ctx);
+ if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
+ 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;
+ unsigned long user_locked, user_lock_limit;
+ struct user_struct *user = current_user();
+ unsigned long locked, lock_limit;
unsigned long vma_size;
unsigned long nr_pages;
- unsigned long locked, lock_limit;
+ long user_extra, extra;
int ret = 0;
- long extra;
if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
return -EINVAL;
if (vma->vm_pgoff != 0)
return -EINVAL;
+ WARN_ON_ONCE(counter->ctx->parent_ctx);
mutex_lock(&counter->mmap_mutex);
if (atomic_inc_not_zero(&counter->mmap_count)) {
if (nr_pages != counter->data->nr_pages)
goto unlock;
}
- extra = nr_pages /* + 1 only account the data pages */;
- extra -= sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
- if (extra < 0)
- extra = 0;
+ user_extra = nr_pages + 1;
+ user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
- locked = vma->vm_mm->locked_vm + extra;
+ /*
+ * Increase the limit linearly with more CPUs:
+ */
+ user_lock_limit *= num_online_cpus();
+
+ 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);
+ atomic_long_add(user_extra, &user->locked_vm);
vma->vm_mm->locked_vm += extra;
counter->data->nr_locked = extra;
unlock:
static int perf_fasync(int fd, struct file *filp, int on)
{
- struct perf_counter *counter = filp->private_data;
struct inode *inode = filp->f_path.dentry->d_inode;
+ struct perf_counter *counter = filp->private_data;
int retval;
mutex_lock(&inode->i_mutex);
struct perf_output_handle {
struct perf_counter *counter;
struct perf_mmap_data *data;
- unsigned int offset;
- unsigned int head;
+ unsigned long head;
+ unsigned long offset;
int nmi;
int overflow;
int locked;
static void perf_output_unlock(struct perf_output_handle *handle)
{
struct perf_mmap_data *data = handle->data;
- int head, cpu;
+ unsigned long head;
+ int cpu;
data->done_head = data->head;
* before we publish the new head, matched by a rmb() in userspace when
* reading this position.
*/
- while ((head = atomic_xchg(&data->done_head, 0)))
+ while ((head = atomic_long_xchg(&data->done_head, 0)))
data->user_page->data_head = head;
/*
/*
* Therefore we have to validate we did not indeed do so.
*/
- if (unlikely(atomic_read(&data->done_head))) {
+ if (unlikely(atomic_long_read(&data->done_head))) {
/*
* Since we had it locked, we can lock it again.
*/
do {
offset = head = atomic_read(&data->head);
head += size;
- } while (atomic_cmpxchg(&data->head, offset, head) != offset);
+ } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
handle->offset = offset;
handle->head = 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);
+ WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
}
#define perf_output_put(handle, x) \
struct perf_counter *counter = handle->counter;
struct perf_mmap_data *data = handle->data;
- int wakeup_events = counter->hw_event.wakeup_events;
+ int wakeup_events = counter->attr.wakeup_events;
if (handle->overflow && wakeup_events) {
int events = atomic_inc_return(&data->events);
rcu_read_unlock();
}
+static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
+{
+ /*
+ * only top level counters have the pid namespace they were created in
+ */
+ if (counter->parent)
+ counter = counter->parent;
+
+ return task_tgid_nr_ns(p, counter->ns);
+}
+
+static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
+{
+ /*
+ * only top level counters have the pid namespace they were created in
+ */
+ if (counter->parent)
+ counter = counter->parent;
+
+ return task_pid_nr_ns(p, counter->ns);
+}
+
static void perf_counter_output(struct perf_counter *counter,
int nmi, struct pt_regs *regs, u64 addr)
{
int ret;
- u64 record_type = counter->hw_event.record_type;
+ u64 sample_type = counter->attr.sample_type;
struct perf_output_handle handle;
struct perf_event_header header;
u64 ip;
u32 pid, tid;
} tid_entry;
struct {
- u64 event;
+ u64 id;
u64 counter;
} group_entry;
struct perf_callchain_entry *callchain = NULL;
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);
- header.type |= PERF_RECORD_IP;
+ if (sample_type & PERF_SAMPLE_IP) {
+ ip = perf_instruction_pointer(regs);
+ header.type |= PERF_SAMPLE_IP;
header.size += sizeof(ip);
}
- if (record_type & PERF_RECORD_TID) {
+ if (sample_type & PERF_SAMPLE_TID) {
/* namespace issues */
- tid_entry.pid = current->group_leader->pid;
- tid_entry.tid = current->pid;
+ tid_entry.pid = perf_counter_pid(counter, current);
+ tid_entry.tid = perf_counter_tid(counter, current);
- header.type |= PERF_RECORD_TID;
+ header.type |= PERF_SAMPLE_TID;
header.size += sizeof(tid_entry);
}
- if (record_type & PERF_RECORD_TIME) {
+ if (sample_type & PERF_SAMPLE_TIME) {
/*
* Maybe do better on x86 and provide cpu_clock_nmi()
*/
time = sched_clock();
- header.type |= PERF_RECORD_TIME;
+ header.type |= PERF_SAMPLE_TIME;
header.size += sizeof(u64);
}
- if (record_type & PERF_RECORD_ADDR) {
- header.type |= PERF_RECORD_ADDR;
+ if (sample_type & PERF_SAMPLE_ADDR) {
+ header.type |= PERF_SAMPLE_ADDR;
header.size += sizeof(u64);
}
- if (record_type & PERF_RECORD_CONFIG) {
- header.type |= PERF_RECORD_CONFIG;
+ if (sample_type & PERF_SAMPLE_CONFIG) {
+ header.type |= PERF_SAMPLE_CONFIG;
header.size += sizeof(u64);
}
- if (record_type & PERF_RECORD_CPU) {
- header.type |= PERF_RECORD_CPU;
+ if (sample_type & PERF_SAMPLE_CPU) {
+ header.type |= PERF_SAMPLE_CPU;
header.size += sizeof(cpu_entry);
cpu_entry.cpu = raw_smp_processor_id();
}
- if (record_type & PERF_RECORD_GROUP) {
- header.type |= PERF_RECORD_GROUP;
+ if (sample_type & PERF_SAMPLE_GROUP) {
+ header.type |= PERF_SAMPLE_GROUP;
header.size += sizeof(u64) +
counter->nr_siblings * sizeof(group_entry);
}
- if (record_type & PERF_RECORD_CALLCHAIN) {
+ if (sample_type & PERF_SAMPLE_CALLCHAIN) {
callchain = perf_callchain(regs);
if (callchain) {
callchain_size = (1 + callchain->nr) * sizeof(u64);
- header.type |= PERF_RECORD_CALLCHAIN;
+ header.type |= PERF_SAMPLE_CALLCHAIN;
header.size += callchain_size;
}
}
perf_output_put(&handle, header);
- if (record_type & PERF_RECORD_IP)
+ if (sample_type & PERF_SAMPLE_IP)
perf_output_put(&handle, ip);
- if (record_type & PERF_RECORD_TID)
+ if (sample_type & PERF_SAMPLE_TID)
perf_output_put(&handle, tid_entry);
- if (record_type & PERF_RECORD_TIME)
+ if (sample_type & PERF_SAMPLE_TIME)
perf_output_put(&handle, time);
- if (record_type & PERF_RECORD_ADDR)
+ if (sample_type & PERF_SAMPLE_ADDR)
perf_output_put(&handle, addr);
- if (record_type & PERF_RECORD_CONFIG)
- perf_output_put(&handle, counter->hw_event.config);
+ if (sample_type & PERF_SAMPLE_CONFIG)
+ perf_output_put(&handle, counter->attr.config);
- if (record_type & PERF_RECORD_CPU)
+ if (sample_type & PERF_SAMPLE_CPU)
perf_output_put(&handle, cpu_entry);
/*
- * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
+ * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
*/
- if (record_type & PERF_RECORD_GROUP) {
+ if (sample_type & PERF_SAMPLE_GROUP) {
struct perf_counter *leader, *sub;
u64 nr = counter->nr_siblings;
if (sub != counter)
sub->pmu->read(sub);
- group_entry.event = sub->hw_event.config;
+ group_entry.id = sub->id;
group_entry.counter = atomic64_read(&sub->count);
perf_output_put(&handle, group_entry);
perf_output_end(&handle);
}
+/*
+ * fork tracking
+ */
+
+struct perf_fork_event {
+ struct task_struct *task;
+
+ struct {
+ struct perf_event_header header;
+
+ u32 pid;
+ u32 ppid;
+ } event;
+};
+
+static void perf_counter_fork_output(struct perf_counter *counter,
+ struct perf_fork_event *fork_event)
+{
+ struct perf_output_handle handle;
+ int size = fork_event->event.header.size;
+ struct task_struct *task = fork_event->task;
+ int ret = perf_output_begin(&handle, counter, size, 0, 0);
+
+ if (ret)
+ return;
+
+ fork_event->event.pid = perf_counter_pid(counter, task);
+ fork_event->event.ppid = perf_counter_pid(counter, task->real_parent);
+
+ perf_output_put(&handle, fork_event->event);
+ perf_output_end(&handle);
+}
+
+static int perf_counter_fork_match(struct perf_counter *counter)
+{
+ if (counter->attr.comm || counter->attr.mmap || counter->attr.munmap)
+ return 1;
+
+ return 0;
+}
+
+static void perf_counter_fork_ctx(struct perf_counter_context *ctx,
+ struct perf_fork_event *fork_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_fork_match(counter))
+ perf_counter_fork_output(counter, fork_event);
+ }
+ rcu_read_unlock();
+}
+
+static void perf_counter_fork_event(struct perf_fork_event *fork_event)
+{
+ struct perf_cpu_context *cpuctx;
+ struct perf_counter_context *ctx;
+
+ cpuctx = &get_cpu_var(perf_cpu_context);
+ perf_counter_fork_ctx(&cpuctx->ctx, fork_event);
+ put_cpu_var(perf_cpu_context);
+
+ rcu_read_lock();
+ /*
+ * doesn't really matter which of the child contexts the
+ * events ends up in.
+ */
+ ctx = rcu_dereference(current->perf_counter_ctxp);
+ if (ctx)
+ perf_counter_fork_ctx(ctx, fork_event);
+ rcu_read_unlock();
+}
+
+void perf_counter_fork(struct task_struct *task)
+{
+ struct perf_fork_event fork_event;
+
+ if (!atomic_read(&nr_comm_counters) &&
+ !atomic_read(&nr_mmap_counters) &&
+ !atomic_read(&nr_munmap_counters))
+ return;
+
+ fork_event = (struct perf_fork_event){
+ .task = task,
+ .event = {
+ .header = {
+ .type = PERF_EVENT_FORK,
+ .size = sizeof(fork_event.event),
+ },
+ },
+ };
+
+ perf_counter_fork_event(&fork_event);
+}
+
/*
* comm tracking
*/
struct perf_comm_event {
- struct task_struct *task;
- char *comm;
+ struct task_struct *task;
+ char *comm;
int comm_size;
struct {
if (ret)
return;
+ comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
+ comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
+
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)
+static int perf_counter_comm_match(struct perf_counter *counter)
{
- if (counter->hw_event.comm &&
- comm_event->event.header.type == PERF_EVENT_COMM)
+ if (counter->attr.comm)
return 1;
return 0;
rcu_read_lock();
list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
- if (perf_counter_comm_match(counter, comm_event))
+ if (perf_counter_comm_match(counter))
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;
+ struct perf_counter_context *ctx;
unsigned int size;
char *comm = comm_event->task->comm;
perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
put_cpu_var(perf_cpu_context);
- perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
+ rcu_read_lock();
+ /*
+ * doesn't really matter which of the child contexts the
+ * events ends up in.
+ */
+ ctx = rcu_dereference(current->perf_counter_ctxp);
+ if (ctx)
+ perf_counter_comm_ctx(ctx, comm_event);
+ rcu_read_unlock();
}
void perf_counter_comm(struct task_struct *task)
{
struct perf_comm_event comm_event;
- if (!atomic_read(&nr_comm_tracking))
+ if (!atomic_read(&nr_comm_counters))
return;
-
+
comm_event = (struct perf_comm_event){
.task = task,
.event = {
.header = { .type = PERF_EVENT_COMM, },
- .pid = task->group_leader->pid,
- .tid = task->pid,
},
};
if (ret)
return;
+ mmap_event->event.pid = perf_counter_pid(counter, current);
+ mmap_event->event.tid = perf_counter_tid(counter, current);
+
perf_output_put(&handle, mmap_event->event);
perf_output_copy(&handle, mmap_event->file_name,
mmap_event->file_size);
static int perf_counter_mmap_match(struct perf_counter *counter,
struct perf_mmap_event *mmap_event)
{
- if (counter->hw_event.mmap &&
+ if (counter->attr.mmap &&
mmap_event->event.header.type == PERF_EVENT_MMAP)
return 1;
- if (counter->hw_event.munmap &&
+ if (counter->attr.munmap &&
mmap_event->event.header.type == PERF_EVENT_MUNMAP)
return 1;
static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
{
struct perf_cpu_context *cpuctx;
+ struct perf_counter_context *ctx;
struct file *file = mmap_event->file;
unsigned int size;
char tmp[16];
perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
put_cpu_var(perf_cpu_context);
- perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
+ rcu_read_lock();
+ /*
+ * doesn't really matter which of the child contexts the
+ * events ends up in.
+ */
+ ctx = rcu_dereference(current->perf_counter_ctxp);
+ if (ctx)
+ perf_counter_mmap_ctx(ctx, mmap_event);
+ rcu_read_unlock();
kfree(buf);
}
{
struct perf_mmap_event mmap_event;
- if (!atomic_read(&nr_mmap_tracking))
+ if (!atomic_read(&nr_mmap_counters))
return;
mmap_event = (struct perf_mmap_event){
.file = file,
.event = {
.header = { .type = PERF_EVENT_MMAP, },
- .pid = current->group_leader->pid,
- .tid = current->pid,
.start = addr,
.len = len,
.pgoff = pgoff,
{
struct perf_mmap_event mmap_event;
- if (!atomic_read(&nr_munmap_tracking))
+ if (!atomic_read(&nr_munmap_counters))
return;
mmap_event = (struct perf_mmap_event){
.file = file,
.event = {
.header = { .type = PERF_EVENT_MUNMAP, },
- .pid = current->group_leader->pid,
- .tid = current->pid,
.start = addr,
.len = len,
.pgoff = pgoff,
perf_counter_mmap_event(&mmap_event);
}
+/*
+ * Log sample_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.sample_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);
+}
+
+/*
+ * IRQ throttle logging
+ */
+
+static void perf_log_throttle(struct perf_counter *counter, int enable)
+{
+ struct perf_output_handle handle;
+ int ret;
+
+ struct {
+ struct perf_event_header header;
+ u64 time;
+ } throttle_event = {
+ .header = {
+ .type = PERF_EVENT_THROTTLE + 1,
+ .misc = 0,
+ .size = sizeof(throttle_event),
+ },
+ .time = sched_clock(),
+ };
+
+ ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
+ if (ret)
+ return;
+
+ perf_output_put(&handle, throttle_event);
+ perf_output_end(&handle);
+}
+
/*
* Generic counter overflow handling.
*/
int nmi, struct pt_regs *regs, u64 addr)
{
int events = atomic_read(&counter->event_limit);
+ int throttle = counter->pmu->unthrottle != NULL;
int ret = 0;
+ if (!throttle) {
+ counter->hw.interrupts++;
+ } else {
+ if (counter->hw.interrupts != MAX_INTERRUPTS) {
+ counter->hw.interrupts++;
+ if (HZ*counter->hw.interrupts > (u64)sysctl_perf_counter_limit) {
+ counter->hw.interrupts = MAX_INTERRUPTS;
+ perf_log_throttle(counter, 0);
+ ret = 1;
+ }
+ } else {
+ /*
+ * Keep re-disabling counters even though on the previous
+ * pass we disabled it - just in case we raced with a
+ * sched-in and the counter got enabled again:
+ */
+ ret = 1;
+ }
+ }
+
/*
* XXX event_limit might not quite work as expected on inherited
* counters
{
struct hw_perf_counter *hwc = &counter->hw;
s64 left = atomic64_read(&hwc->period_left);
- s64 period = hwc->irq_period;
+ s64 period = hwc->sample_period;
if (unlikely(left <= -period)) {
left = period;
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->pmu->read(counter);
* In case we exclude kernel IPs or are somehow not in interrupt
* context, provide the next best thing, the user IP.
*/
- if ((counter->hw_event.exclude_kernel || !regs) &&
- !counter->hw_event.exclude_user)
+ if ((counter->attr.exclude_kernel || !regs) &&
+ !counter->attr.exclude_user)
regs = task_pt_regs(current);
if (regs) {
ret = HRTIMER_NORESTART;
}
- hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
+ period = max_t(u64, 10000, counter->hw.sample_period);
+ hrtimer_forward_now(hrtimer, ns_to_ktime(period));
return ret;
}
}
+static int perf_swcounter_is_counting(struct perf_counter *counter)
+{
+ struct perf_counter_context *ctx;
+ unsigned long flags;
+ int count;
+
+ if (counter->state == PERF_COUNTER_STATE_ACTIVE)
+ return 1;
+
+ if (counter->state != PERF_COUNTER_STATE_INACTIVE)
+ return 0;
+
+ /*
+ * If the counter is inactive, it could be just because
+ * its task is scheduled out, or because it's in a group
+ * which could not go on the PMU. We want to count in
+ * the first case but not the second. If the context is
+ * currently active then an inactive software counter must
+ * be the second case. If it's not currently active then
+ * we need to know whether the counter was active when the
+ * context was last active, which we can determine by
+ * comparing counter->tstamp_stopped with ctx->time.
+ *
+ * We are within an RCU read-side critical section,
+ * which protects the existence of *ctx.
+ */
+ ctx = counter->ctx;
+ spin_lock_irqsave(&ctx->lock, flags);
+ count = 1;
+ /* Re-check state now we have the lock */
+ if (counter->state < PERF_COUNTER_STATE_INACTIVE ||
+ counter->ctx->is_active ||
+ counter->tstamp_stopped < ctx->time)
+ count = 0;
+ spin_unlock_irqrestore(&ctx->lock, flags);
+ return count;
+}
+
static int perf_swcounter_match(struct perf_counter *counter,
enum perf_event_types type,
u32 event, struct pt_regs *regs)
{
- if (counter->state != PERF_COUNTER_STATE_ACTIVE)
- return 0;
+ u64 event_config;
- if (perf_event_raw(&counter->hw_event))
- return 0;
+ event_config = ((u64) type << PERF_COUNTER_TYPE_SHIFT) | event;
- if (perf_event_type(&counter->hw_event) != type)
+ if (!perf_swcounter_is_counting(counter))
return 0;
- if (perf_event_id(&counter->hw_event) != event)
+ if (counter->attr.config != event_config)
return 0;
- if (counter->hw_event.exclude_user && user_mode(regs))
- return 0;
+ if (regs) {
+ if (counter->attr.exclude_user && user_mode(regs))
+ return 0;
- if (counter->hw_event.exclude_kernel && !user_mode(regs))
- return 0;
+ if (counter->attr.exclude_kernel && !user_mode(regs))
+ return 0;
+ }
return 1;
}
int nmi, struct pt_regs *regs, u64 addr)
{
int neg = atomic64_add_negative(nr, &counter->hw.count);
- if (counter->hw.irq_period && !neg)
+
+ if (counter->hw.sample_period && !neg && regs)
perf_swcounter_overflow(counter, nmi, regs, addr);
}
{
struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
int *recursion = perf_swcounter_recursion_context(cpuctx);
+ struct perf_counter_context *ctx;
if (*recursion)
goto out;
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, addr);
- }
+ rcu_read_lock();
+ /*
+ * doesn't really matter which of the child contexts the
+ * events ends up in.
+ */
+ ctx = rcu_dereference(current->perf_counter_ctxp);
+ if (ctx)
+ perf_swcounter_ctx_event(ctx, type, event, nr, nmi, regs, addr);
+ rcu_read_unlock();
barrier();
(*recursion)--;
atomic64_set(&hwc->prev_count, cpu_clock(cpu));
hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
hwc->hrtimer.function = perf_swcounter_hrtimer;
- if (hwc->irq_period) {
+ if (hwc->sample_period) {
+ u64 period = max_t(u64, 10000, hwc->sample_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.sample_period)
+ hrtimer_cancel(&counter->hw.hrtimer);
cpu_clock_perf_counter_update(counter);
}
atomic64_set(&hwc->prev_count, now);
hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
hwc->hrtimer.function = perf_swcounter_hrtimer;
- if (hwc->irq_period) {
+ if (hwc->sample_period) {
+ u64 period = max_t(u64, 10000, hwc->sample_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.sample_period)
+ hrtimer_cancel(&counter->hw.hrtimer);
task_clock_perf_counter_update(counter, counter->ctx->time);
}
/*
* Software counter: cpu migrations
*/
-
-static inline u64 get_cpu_migrations(struct perf_counter *counter)
-{
- struct task_struct *curr = counter->ctx->task;
-
- if (curr)
- return curr->se.nr_migrations;
- return cpu_nr_migrations(smp_processor_id());
-}
-
-static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
-{
- u64 prev, now;
- s64 delta;
-
- prev = atomic64_read(&counter->hw.prev_count);
- now = get_cpu_migrations(counter);
-
- atomic64_set(&counter->hw.prev_count, now);
-
- delta = now - prev;
-
- atomic64_add(delta, &counter->count);
-}
-
-static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
+void perf_counter_task_migration(struct task_struct *task, int cpu)
{
- cpu_migrations_perf_counter_update(counter);
-}
+ struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
+ struct perf_counter_context *ctx;
-static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
-{
- if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
- atomic64_set(&counter->hw.prev_count,
- get_cpu_migrations(counter));
- return 0;
-}
+ perf_swcounter_ctx_event(&cpuctx->ctx, PERF_TYPE_SOFTWARE,
+ PERF_COUNT_CPU_MIGRATIONS,
+ 1, 1, NULL, 0);
-static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
-{
- cpu_migrations_perf_counter_update(counter);
+ ctx = perf_pin_task_context(task);
+ if (ctx) {
+ perf_swcounter_ctx_event(ctx, PERF_TYPE_SOFTWARE,
+ PERF_COUNT_CPU_MIGRATIONS,
+ 1, 1, NULL, 0);
+ perf_unpin_context(ctx);
+ }
}
-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,
-};
-
#ifdef CONFIG_EVENT_PROFILE
void perf_tpcounter_event(int event_id)
{
static void tp_perf_counter_destroy(struct perf_counter *counter)
{
- ftrace_profile_disable(perf_event_id(&counter->hw_event));
+ ftrace_profile_disable(perf_event_id(&counter->attr));
}
static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
{
- int event_id = perf_event_id(&counter->hw_event);
+ int event_id = perf_event_id(&counter->attr);
int ret;
ret = ftrace_profile_enable(event_id);
return NULL;
counter->destroy = tp_perf_counter_destroy;
- counter->hw.irq_period = counter->hw_event.irq_period;
+ counter->hw.sample_period = counter->attr.sample_period;
return &perf_ops_generic;
}
static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
{
- struct perf_counter_hw_event *hw_event = &counter->hw_event;
const struct pmu *pmu = NULL;
- struct hw_perf_counter *hwc = &counter->hw;
/*
* Software counters (currently) can't in general distinguish
* to be kernel events, and page faults are never hypervisor
* events.
*/
- switch (perf_event_id(&counter->hw_event)) {
+ switch (perf_event_id(&counter->attr)) {
case PERF_COUNT_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:
/*
else
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:
- pmu = &perf_ops_generic;
- break;
case PERF_COUNT_CPU_MIGRATIONS:
- if (!counter->hw_event.exclude_kernel)
- pmu = &perf_ops_cpu_migrations;
+ pmu = &perf_ops_generic;
break;
}
- if (pmu)
- hwc->irq_period = hw_event->irq_period;
-
return pmu;
}
* Allocate and initialize a counter structure
*/
static struct perf_counter *
-perf_counter_alloc(struct perf_counter_hw_event *hw_event,
+perf_counter_alloc(struct perf_counter_attr *attr,
int cpu,
struct perf_counter_context *ctx,
struct perf_counter *group_leader,
{
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->attr = *attr;
+ counter->group_leader = group_leader;
+ counter->pmu = NULL;
+ counter->ctx = ctx;
+ counter->oncpu = -1;
- counter->cpu = cpu;
- counter->hw_event = *hw_event;
- counter->group_leader = group_leader;
- counter->pmu = NULL;
- counter->ctx = ctx;
+ counter->ns = get_pid_ns(current->nsproxy->pid_ns);
+ counter->id = atomic64_inc_return(&perf_counter_id);
- counter->state = PERF_COUNTER_STATE_INACTIVE;
- if (hw_event->disabled)
+ counter->state = PERF_COUNTER_STATE_INACTIVE;
+
+ if (attr->disabled)
counter->state = PERF_COUNTER_STATE_OFF;
pmu = NULL;
+ hwc = &counter->hw;
+ if (attr->freq && attr->sample_freq)
+ hwc->sample_period = div64_u64(TICK_NSEC, attr->sample_freq);
+ else
+ hwc->sample_period = attr->sample_period;
+
/*
- * we currently do not support PERF_RECORD_GROUP on inherited counters
+ * we currently do not support PERF_SAMPLE_GROUP on inherited counters
*/
- if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
+ if (attr->inherit && (attr->sample_type & PERF_SAMPLE_GROUP))
goto done;
- if (perf_event_raw(hw_event)) {
+ if (perf_event_raw(attr)) {
pmu = hw_perf_counter_init(counter);
goto done;
}
- switch (perf_event_type(hw_event)) {
+ switch (perf_event_type(attr)) {
case PERF_TYPE_HARDWARE:
pmu = hw_perf_counter_init(counter);
break;
err = PTR_ERR(pmu);
if (err) {
+ if (counter->ns)
+ put_pid_ns(counter->ns);
kfree(counter);
return ERR_PTR(err);
}
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);
+ if (counter->attr.mmap)
+ atomic_inc(&nr_mmap_counters);
+ if (counter->attr.munmap)
+ atomic_inc(&nr_munmap_counters);
+ if (counter->attr.comm)
+ atomic_inc(&nr_comm_counters);
return counter;
}
/**
* sys_perf_counter_open - open a performance counter, associate it to a task/cpu
*
- * @hw_event_uptr: event type attributes for monitoring/sampling
+ * @attr_uptr: event type attributes for monitoring/sampling
* @pid: target pid
* @cpu: target cpu
* @group_fd: group leader counter fd
*/
SYSCALL_DEFINE5(perf_counter_open,
- const struct perf_counter_hw_event __user *, hw_event_uptr,
+ const struct perf_counter_attr __user *, attr_uptr,
pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
{
struct perf_counter *counter, *group_leader;
- struct perf_counter_hw_event hw_event;
+ struct perf_counter_attr attr;
struct perf_counter_context *ctx;
struct file *counter_file = NULL;
struct file *group_file = NULL;
if (flags)
return -EINVAL;
- if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
+ if (copy_from_user(&attr, attr_uptr, sizeof(attr)) != 0)
return -EFAULT;
/*
/*
* Only a group leader can be exclusive or pinned
*/
- if (hw_event.exclusive || hw_event.pinned)
+ if (attr.exclusive || attr.pinned)
goto err_put_context;
}
- counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
+ counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
GFP_KERNEL);
ret = PTR_ERR(counter);
if (IS_ERR(counter))
goto err_free_put_context;
counter->filp = counter_file;
+ WARN_ON_ONCE(ctx->parent_ctx);
mutex_lock(&ctx->mutex);
perf_install_in_context(ctx, counter, cpu);
+ ++ctx->generation;
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:
kfree(counter);
err_put_context:
- put_context(ctx);
+ put_ctx(ctx);
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 (parent_counter->parent)
parent_counter = parent_counter->parent;
- child_counter = perf_counter_alloc(&parent_counter->hw_event,
+ child_counter = perf_counter_alloc(&parent_counter->attr,
parent_counter->cpu, child_ctx,
group_leader, GFP_KERNEL);
if (IS_ERR(child_counter))
return child_counter;
+ get_ctx(child_ctx);
+
+ /*
+ * Make the child state follow the state of the parent counter,
+ * not its attr.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;
/*
* inherit into child's child as well:
*/
- child_counter->hw_event.inherit = 1;
+ child_counter->attr.inherit = 1;
/*
* Get a reference to the parent filp - we will fput it
/*
* Link this into the parent counter's child list
*/
- mutex_lock(&parent_counter->mutex);
+ WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
+ 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);
+ WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
+ 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
}
static void
-__perf_counter_exit_task(struct task_struct *child,
- struct perf_counter *child_counter,
+__perf_counter_exit_task(struct perf_counter *child_counter,
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);
}
}
/*
* When a child task exits, feed back counter values to parent counters.
- *
- * Note: we may be running in child context, but the PID is not hashed
- * anymore so new counters will not be added.
*/
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;
-
- if (likely(!child_ctx->nr_counters))
+ if (likely(!child->perf_counter_ctxp))
return;
+ local_irq_save(flags);
+ /*
+ * We can't reschedule here because interrupts are disabled,
+ * and either child is current or it is a task that can't be
+ * scheduled, so we are now safe from rescheduling changing
+ * our context.
+ */
+ child_ctx = child->perf_counter_ctxp;
+ __perf_counter_task_sched_out(child_ctx);
+
+ /*
+ * Take the context lock here so that if find_get_context is
+ * reading child->perf_counter_ctxp, we wait until it has
+ * incremented the context's refcount before we do put_ctx below.
+ */
+ spin_lock(&child_ctx->lock);
+ child->perf_counter_ctxp = NULL;
+ if (child_ctx->parent_ctx) {
+ /*
+ * This context is a clone; unclone it so it can't get
+ * swapped to another process while we're removing all
+ * the counters from it.
+ */
+ put_ctx(child_ctx->parent_ctx);
+ child_ctx->parent_ctx = NULL;
+ }
+ spin_unlock(&child_ctx->lock);
+ 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);
+ __perf_counter_exit_task(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);
+}
+
+/*
+ * free an unexposed, unused context as created by inheritance by
+ * init_task below, used by fork() in case of fail.
+ */
+void perf_counter_free_task(struct task_struct *task)
+{
+ struct perf_counter_context *ctx = task->perf_counter_ctxp;
+ struct perf_counter *counter, *tmp;
+
+ if (!ctx)
+ return;
+
+ mutex_lock(&ctx->mutex);
+again:
+ list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
+ struct perf_counter *parent = counter->parent;
+
+ if (WARN_ON_ONCE(!parent))
+ continue;
+
+ mutex_lock(&parent->child_mutex);
+ list_del_init(&counter->child_list);
+ mutex_unlock(&parent->child_mutex);
+
+ fput(parent->filp);
+
+ list_del_counter(counter, ctx);
+ free_counter(counter);
+ }
+
+ if (!list_empty(&ctx->counter_list))
+ goto again;
+
+ mutex_unlock(&ctx->mutex);
+
+ put_ctx(ctx);
}
/*
* Initialize the perf_counter context in task_struct
*/
-void perf_counter_init_task(struct task_struct *child)
+int perf_counter_init_task(struct task_struct *child)
{
struct perf_counter_context *child_ctx, *parent_ctx;
+ struct perf_counter_context *cloned_ctx;
struct perf_counter *counter;
struct task_struct *parent = current;
+ int inherited_all = 1;
+ int ret = 0;
- 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);
+
+ if (likely(!parent->perf_counter_ctxp))
+ return 0;
/*
* 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))
- return;
+ child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
+ if (!child_ctx)
+ return -ENOMEM;
+
+ __perf_counter_init_context(child_ctx, child);
+ child->perf_counter_ctxp = child_ctx;
+ get_task_struct(child);
+
+ /*
+ * If the parent's context is a clone, pin it so it won't get
+ * swapped under us.
+ */
+ parent_ctx = perf_pin_task_context(parent);
+
+ /*
+ * No need to check if parent_ctx != NULL here; since we saw
+ * it non-NULL earlier, the only reason for it to become NULL
+ * is if we exit, and since we're currently in the middle of
+ * a fork we can't be exiting at the same time.
+ */
/*
* Lock the parent list. No need to lock the child - not PID
* 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->attr.inherit) {
+ inherited_all = 0;
continue;
+ }
- if (inherit_group(counter, parent,
- parent_ctx, child, child_ctx))
+ ret = inherit_group(counter, parent, parent_ctx,
+ child, child_ctx);
+ if (ret) {
+ 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.
+ * Note that if the parent is a clone, it could get
+ * uncloned at any point, but that doesn't matter
+ * because the list of counters and the generation
+ * count can't have changed since we took the mutex.
+ */
+ cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
+ if (cloned_ctx) {
+ child_ctx->parent_ctx = cloned_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);
+
+ perf_unpin_context(parent_ctx);
+
+ return ret;
}
static void __cpuinit perf_counter_init_cpu(int cpu)
return NOTIFY_OK;
}
+/*
+ * This has to have a higher priority than migration_notifier in sched.c.
+ */
static struct notifier_block __cpuinitdata perf_cpu_nb = {
.notifier_call = perf_cpu_notify,
+ .priority = 20,
};
void __init perf_counter_init(void)