Compile tested only, depends on the cgroup::exit patch
---
Subject: perf: Add cgroup support
From: Stephane Eranian <[email protected]>
Date: Mon Feb 07 17:02:25 CET 2011
This kernel patch adds the ability to filter monitoring based on
container groups (cgroups). This is for use in per-cpu mode only.
The cgroup to monitor is passed as a file descriptor in the pid
argument to the syscall. The file descriptor must be opened to
the cgroup name in the cgroup filesystem. For instance, if the
cgroup name is foo and cgroupfs is mounted in /cgroup, then the
file descriptor is opened to /cgroup/foo. Cgroup mode is
activated by passing PERF_FLAG_PID_CGROUP in the flags argument
to the syscall.
For instance to measure in cgroup foo on CPU1 assuming
cgroupfs is mounted under /cgroup:
struct perf_event_attr attr;
int cgroup_fd, fd;
cgroup_fd = open("/cgroup/foo", O_RDONLY);
fd = perf_event_open(&attr, cgroup_fd, 1, -1, PERF_FLAG_PID_CGROUP);
close(cgroup_fd);
Signed-off-by: Stephane Eranian <[email protected]>
[ added perf_cgroup_{exit,attach} ]
Signed-off-by: Peter Zijlstra <[email protected]>
LKML-Reference: <new-submission>
---
include/linux/cgroup.h | 1
include/linux/cgroup_subsys.h | 4
include/linux/perf_event.h | 33 +-
init/Kconfig | 10
kernel/cgroup.c | 23 +
kernel/perf_event.c | 641 +++++++++++++++++++++++++++++++++++++++---
6 files changed, 665 insertions(+), 47 deletions(-)
Index: linux-2.6/include/linux/cgroup.h
===================================================================
--- linux-2.6.orig/include/linux/cgroup.h
+++ linux-2.6/include/linux/cgroup.h
@@ -627,6 +627,7 @@ bool css_is_ancestor(struct cgroup_subsy
/* Get id and depth of css */
unsigned short css_id(struct cgroup_subsys_state *css);
unsigned short css_depth(struct cgroup_subsys_state *css);
+struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id);
#else /* !CONFIG_CGROUPS */
Index: linux-2.6/include/linux/cgroup_subsys.h
===================================================================
--- linux-2.6.orig/include/linux/cgroup_subsys.h
+++ linux-2.6/include/linux/cgroup_subsys.h
@@ -65,4 +65,8 @@ SUBSYS(net_cls)
SUBSYS(blkio)
#endif
+#ifdef CONFIG_CGROUP_PERF
+SUBSYS(perf)
+#endif
+
/* */
Index: linux-2.6/include/linux/perf_event.h
===================================================================
--- linux-2.6.orig/include/linux/perf_event.h
+++ linux-2.6/include/linux/perf_event.h
@@ -464,6 +464,7 @@ enum perf_callchain_context {
#define PERF_FLAG_FD_NO_GROUP (1U << 0)
#define PERF_FLAG_FD_OUTPUT (1U << 1)
+#define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */
#ifdef __KERNEL__
/*
@@ -471,6 +472,7 @@ enum perf_callchain_context {
*/
#ifdef CONFIG_PERF_EVENTS
+# include <linux/cgroup.h>
# include <asm/perf_event.h>
# include <asm/local64.h>
#endif
@@ -716,6 +718,22 @@ struct swevent_hlist {
#define PERF_ATTACH_GROUP 0x02
#define PERF_ATTACH_TASK 0x04
+#ifdef CONFIG_CGROUP_PERF
+/*
+ * perf_cgroup_info keeps track of time_enabled for a cgroup.
+ * This is a per-cpu dynamically allocated data structure.
+ */
+struct perf_cgroup_info {
+ u64 time;
+ u64 timestamp;
+};
+
+struct perf_cgroup {
+ struct cgroup_subsys_state css;
+ struct perf_cgroup_info *info; /* timing info, one per cpu */
+};
+#endif
+
/**
* struct perf_event - performance event kernel representation:
*/
@@ -832,6 +850,11 @@ struct perf_event {
struct event_filter *filter;
#endif
+#ifdef CONFIG_CGROUP_PERF
+ struct perf_cgroup *cgrp; /* cgroup event is attach to */
+ int cgrp_defer_enabled;
+#endif
+
#endif /* CONFIG_PERF_EVENTS */
};
@@ -886,6 +909,7 @@ struct perf_event_context {
u64 generation;
int pin_count;
struct rcu_head rcu_head;
+ int nr_cgroups; /* cgroup events present */
};
/*
@@ -905,6 +929,9 @@ struct perf_cpu_context {
struct list_head rotation_list;
int jiffies_interval;
struct pmu *active_pmu;
+#ifdef CONFIG_CGROUP_PERF
+ struct perf_cgroup *cgrp;
+#endif
};
struct perf_output_handle {
@@ -1040,11 +1067,11 @@ perf_sw_event(u32 event_id, u64 nr, int
__perf_sw_event(event_id, nr, nmi, regs, addr);
}
-extern atomic_t perf_task_events;
+extern atomic_t perf_sched_events;
static inline void perf_event_task_sched_in(struct task_struct *task)
{
- COND_STMT(&perf_task_events, __perf_event_task_sched_in(task));
+ COND_STMT(&perf_sched_events, __perf_event_task_sched_in(task));
}
static inline
@@ -1052,7 +1079,7 @@ void perf_event_task_sched_out(struct ta
{
perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
- COND_STMT(&perf_task_events, __perf_event_task_sched_out(task, next));
+ COND_STMT(&perf_sched_events, __perf_event_task_sched_out(task, next));
}
extern void perf_event_mmap(struct vm_area_struct *vma);
Index: linux-2.6/init/Kconfig
===================================================================
--- linux-2.6.orig/init/Kconfig
+++ linux-2.6/init/Kconfig
@@ -683,6 +683,16 @@ config CGROUP_MEM_RES_CTLR_SWAP_ENABLED
select this option (if, for some reason, they need to disable it
then noswapaccount does the trick).
+config CGROUP_PERF
+ bool "Enable perf_event per-cpu per-container group (cgroup) monitoring"
+ depends on PERF_EVENTS && CGROUPS
+ help
+ This option extends the per-cpu mode to restrict monitoring to
+ threads which belong to the cgroup specificied and run on the
+ designated cpu.
+
+ Say N if unsure.
+
menuconfig CGROUP_SCHED
bool "Group CPU scheduler"
depends on EXPERIMENTAL
Index: linux-2.6/kernel/cgroup.c
===================================================================
--- linux-2.6.orig/kernel/cgroup.c
+++ linux-2.6/kernel/cgroup.c
@@ -4822,6 +4822,29 @@ css_get_next(struct cgroup_subsys *ss, i
return ret;
}
+/*
+ * get corresponding css from file open on cgroupfs directory
+ */
+struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
+{
+ struct cgroup *cgrp;
+ struct inode *inode;
+ struct cgroup_subsys_state *css;
+
+ inode = f->f_dentry->d_inode;
+ /* check in cgroup filesystem dir */
+ if (inode->i_op != &cgroup_dir_inode_operations)
+ return ERR_PTR(-EBADF);
+
+ if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
+ return ERR_PTR(-EINVAL);
+
+ /* get cgroup */
+ cgrp = __d_cgrp(f->f_dentry);
+ css = cgrp->subsys[id];
+ return css ? css : ERR_PTR(-ENOENT);
+}
+
#ifdef CONFIG_CGROUP_DEBUG
static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
struct cgroup *cont)
Index: linux-2.6/kernel/perf_event.c
===================================================================
--- linux-2.6.orig/kernel/perf_event.c
+++ linux-2.6/kernel/perf_event.c
@@ -111,13 +111,23 @@ static int cpu_function_call(int cpu, in
return data.ret;
}
+#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
+ PERF_FLAG_FD_OUTPUT |\
+ PERF_FLAG_PID_CGROUP)
+
enum event_type_t {
EVENT_FLEXIBLE = 0x1,
EVENT_PINNED = 0x2,
EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
};
-atomic_t perf_task_events __read_mostly;
+/*
+ * perf_sched_events : >0 events exist
+ * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
+ */
+atomic_t perf_sched_events __read_mostly;
+static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
+
static atomic_t nr_mmap_events __read_mostly;
static atomic_t nr_comm_events __read_mostly;
static atomic_t nr_task_events __read_mostly;
@@ -148,7 +158,11 @@ static void cpu_ctx_sched_out(struct per
enum event_type_t event_type);
static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
- enum event_type_t event_type);
+ enum event_type_t event_type,
+ struct task_struct *task, int cgrp_sw);
+
+static void update_context_time(struct perf_event_context *ctx);
+static u64 perf_event_time(struct perf_event *event);
void __weak perf_event_print_debug(void) { }
@@ -162,6 +176,315 @@ static inline u64 perf_clock(void)
return local_clock();
}
+#ifdef CONFIG_CGROUP_PERF
+
+static inline struct perf_cgroup *
+perf_cgroup_from_task(struct task_struct *task)
+{
+ return container_of(task_subsys_state(task, perf_subsys_id),
+ struct perf_cgroup, css);
+}
+
+static inline bool
+perf_cgroup_match(struct perf_event *event, struct task_struct *task)
+{
+ struct perf_cgroup *cgrp = NULL;
+ if (task)
+ cgrp = perf_cgroup_from_task(task);
+ return !event->cgrp || event->cgrp == cgrp;
+}
+
+static inline void perf_get_cgroup(struct perf_event *event)
+{
+ css_get(&event->cgrp->css);
+}
+
+static inline void perf_put_cgroup(struct perf_event *event)
+{
+ css_put(&event->cgrp->css);
+}
+
+static inline void perf_detach_cgroup(struct perf_event *event)
+{
+ perf_put_cgroup(event);
+ event->cgrp = NULL;
+}
+
+static inline int is_cgroup_event(struct perf_event *event)
+{
+ return event->cgrp != NULL;
+}
+
+static inline u64 perf_cgroup_event_time(struct perf_event *event)
+{
+ struct perf_cgroup_info *t;
+
+ t = per_cpu_ptr(event->cgrp->info, event->cpu);
+ return t->time;
+}
+
+static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
+{
+ struct perf_cgroup_info *info;
+ u64 now;
+
+ now = perf_clock();
+
+ info = this_cpu_ptr(cgrp->info);
+
+ info->time += now - info->timestamp;
+ info->timestamp = now;
+}
+
+static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
+{
+ struct perf_cgroup *cgrp_out = cpuctx->cgrp;
+ if (cgrp_out)
+ __update_cgrp_time(cgrp_out);
+}
+
+static inline void update_cgrp_time_from_event(struct perf_event *event)
+{
+ struct perf_cgroup *cgrp = perf_cgroup_from_task(current);
+ /*
+ * do not update time when cgroup is not active
+ */
+ if (!event->cgrp || cgrp != event->cgrp)
+ return;
+
+ __update_cgrp_time(event->cgrp);
+}
+
+static inline void
+perf_cgroup_set_timestamp(struct task_struct *task, u64 now)
+{
+ struct perf_cgroup *cgrp;
+ struct perf_cgroup_info *info;
+
+ if (!task)
+ return;
+
+ cgrp = perf_cgroup_from_task(task);
+ info = per_cpu_ptr(cgrp->info, smp_processor_id());
+ info->timestamp = now;
+}
+
+#define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
+#define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
+
+/*
+ * reschedule events based on the cgroup constraint of task.
+ *
+ * mode SWOUT : schedule out everything
+ * mode SWIN : schedule in based on cgroup for next
+ */
+void perf_cgroup_switch(struct task_struct *task, int mode)
+{
+ struct perf_cpu_context *cpuctx;
+ struct pmu *pmu;
+ unsigned long flags;
+
+ /*
+ * disable interrupts to avoid geting nr_cgroup
+ * changes via __perf_event_disable(). Also
+ * avoids preemption.
+ */
+ local_irq_save(flags);
+
+ /*
+ * we reschedule only in the presence of cgroup
+ * constrained events.
+ */
+ rcu_read_lock();
+
+ list_for_each_entry_rcu(pmu, &pmus, entry) {
+
+ cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
+
+ perf_pmu_disable(cpuctx->ctx.pmu);
+
+ /*
+ * perf_cgroup_events says at least one
+ * context on this CPU has cgroup events.
+ *
+ * ctx->nr_cgroups reports the number of cgroup
+ * events for a context.
+ */
+ if (cpuctx->ctx.nr_cgroups > 0) {
+
+ if (mode & PERF_CGROUP_SWOUT)
+ cpu_ctx_sched_out(cpuctx, EVENT_ALL);
+
+ if (mode & PERF_CGROUP_SWIN) {
+ cpu_ctx_sched_in(cpuctx, EVENT_ALL, task, 1);
+ cpuctx->cgrp = perf_cgroup_from_task(task);
+ }
+ }
+
+ perf_pmu_enable(cpuctx->ctx.pmu);
+ }
+
+ rcu_read_unlock();
+
+ local_irq_restore(flags);
+}
+
+static inline void perf_cgroup_sched_out(struct task_struct *task)
+{
+ perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
+}
+
+static inline void perf_cgroup_sched_in(struct task_struct *task)
+{
+ perf_cgroup_switch(task, PERF_CGROUP_SWIN);
+}
+
+static inline int perf_cgroup_connect(int fd, struct perf_event *event,
+ struct perf_event_attr *attr,
+ struct perf_event *group_leader)
+{
+ struct perf_cgroup *cgrp;
+ struct cgroup_subsys_state *css;
+ struct file *file;
+ int ret = 0, fput_needed;
+
+ file = fget_light(fd, &fput_needed);
+ if (!file)
+ return -EBADF;
+
+ css = cgroup_css_from_dir(file, perf_subsys_id);
+ if (IS_ERR(css))
+ return PTR_ERR(css);
+
+ cgrp = container_of(css, struct perf_cgroup, css);
+ event->cgrp = cgrp;
+
+ /*
+ * all events in a group must monitor
+ * the same cgroup because a task belongs
+ * to only one perf cgroup at a time
+ */
+ if (group_leader && group_leader->cgrp != cgrp) {
+ perf_detach_cgroup(event);
+ ret = -EINVAL;
+ } else {
+ /* must be done before we fput() the file */
+ perf_get_cgroup(event);
+ }
+ fput_light(file, fput_needed);
+ return ret;
+}
+
+static inline void
+perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
+{
+ struct perf_cgroup_info *t;
+ t = per_cpu_ptr(event->cgrp->info, event->cpu);
+ event->shadow_ctx_time = now - t->timestamp;
+}
+
+static inline void
+perf_cgroup_defer_enabled(struct perf_event *event, struct task_struct *task)
+{
+ /*
+ * when the current task's perf cgroup does not match
+ * the event's, we need to remember to call the
+ * perf_mark_enable() function the first time a task with
+ * a matching perf cgroup is scheduled in.
+ */
+ if (is_cgroup_event(event) && !perf_cgroup_match(event, task))
+ event->cgrp_defer_enabled = 1;
+}
+
+static inline void
+perf_cgroup_mark_enabled(struct perf_event *event,
+ struct perf_event_context *ctx)
+{
+ struct perf_event *sub;
+ u64 tstamp = perf_event_time(event);
+
+ if (!event->cgrp_defer_enabled)
+ return;
+
+ event->cgrp_defer_enabled = 0;
+
+ event->tstamp_enabled = tstamp - event->total_time_enabled;
+ list_for_each_entry(sub, &event->sibling_list, group_entry) {
+ if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
+ sub->tstamp_enabled = tstamp - sub->total_time_enabled;
+ sub->cgrp_defer_enabled = 0;
+ }
+ }
+}
+#else /* !CONFIG_CGROUP_PERF */
+
+static inline bool
+perf_cgroup_match(struct perf_event *event, struct task_struct *task)
+{
+ return true;
+}
+
+static inline void perf_detach_cgroup(struct perf_event *event)
+{}
+
+static inline int is_cgroup_event(struct perf_event *event)
+{
+ return 0;
+}
+
+static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
+{
+ return 0;
+}
+
+static inline void update_cgrp_time_from_event(struct perf_event *event)
+{}
+
+static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
+{}
+
+static inline void perf_cgroup_sched_out(struct task_struct *task)
+{
+}
+
+static inline void perf_cgroup_sched_in(struct task_struct *task)
+{
+}
+
+static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
+ struct perf_event_attr *attr,
+ struct perf_event *group_leader)
+{
+ return -EINVAL;
+}
+
+static inline void
+perf_cgroup_set_timestamp(struct task_struct *task, u64 now)
+{}
+
+void
+perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
+{}
+
+static inline void
+perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
+{}
+
+static inline u64 perf_cgroup_event_time(struct perf_event *event)
+{
+ return 0;
+}
+
+static inline void
+perf_cgroup_defer_enabled(struct perf_event *event, struct task_struct *task)
+{}
+
+static inline void
+perf_cgroup_mark_enabled(struct perf_event *event,
+ struct perf_event_context *ctx)
+{}
+#endif
+
void perf_pmu_disable(struct pmu *pmu)
{
int *count = this_cpu_ptr(pmu->pmu_disable_count);
@@ -343,6 +666,10 @@ static void update_context_time(struct p
static u64 perf_event_time(struct perf_event *event)
{
struct perf_event_context *ctx = event->ctx;
+
+ if (is_cgroup_event(event))
+ return perf_cgroup_event_time(event);
+
return ctx ? ctx->time : 0;
}
@@ -357,9 +684,20 @@ static void update_event_times(struct pe
if (event->state < PERF_EVENT_STATE_INACTIVE ||
event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
return;
-
- if (ctx->is_active)
+ /*
+ * in cgroup mode, time_enabled represents
+ * the time the event was enabled AND active
+ * tasks were in the monitored cgroup. This is
+ * independent of the activity of the context as
+ * there may be a mix of cgroup and non-cgroup events.
+ *
+ * That is why we treat cgroup events differently
+ * here.
+ */
+ if (is_cgroup_event(event))
run_end = perf_event_time(event);
+ else if (ctx->is_active)
+ run_end = ctx->time;
else
run_end = event->tstamp_stopped;
@@ -371,6 +709,7 @@ static void update_event_times(struct pe
run_end = perf_event_time(event);
event->total_time_running = run_end - event->tstamp_running;
+
}
/*
@@ -419,6 +758,17 @@ list_add_event(struct perf_event *event,
list_add_tail(&event->group_entry, list);
}
+ if (is_cgroup_event(event)) {
+ ctx->nr_cgroups++;
+ /*
+ * one more event:
+ * - that has cgroup constraint on event->cpu
+ * - that may need work on context switch
+ */
+ atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
+ jump_label_inc(&perf_sched_events);
+ }
+
list_add_rcu(&event->event_entry, &ctx->event_list);
if (!ctx->nr_events)
perf_pmu_rotate_start(ctx->pmu);
@@ -545,6 +895,12 @@ list_del_event(struct perf_event *event,
event->attach_state &= ~PERF_ATTACH_CONTEXT;
+ if (is_cgroup_event(event)) {
+ ctx->nr_cgroups--;
+ atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
+ jump_label_dec(&perf_sched_events);
+ }
+
ctx->nr_events--;
if (event->attr.inherit_stat)
ctx->nr_stat--;
@@ -614,9 +970,10 @@ static void perf_group_detach(struct per
}
static inline int
-event_filter_match(struct perf_event *event)
+event_filter_match(struct perf_event *event, struct task_struct *task)
{
- return event->cpu == -1 || event->cpu == smp_processor_id();
+ return (event->cpu == -1 || event->cpu == smp_processor_id())
+ && perf_cgroup_match(event, task);
}
static void
@@ -633,8 +990,8 @@ event_sched_out(struct perf_event *event
* via read() for time_enabled, time_running:
*/
if (event->state == PERF_EVENT_STATE_INACTIVE
- && !event_filter_match(event)) {
- delta = ctx->time - event->tstamp_stopped;
+ && !event_filter_match(event, current)) {
+ delta = tstamp - event->tstamp_stopped;
event->tstamp_running += delta;
event->tstamp_stopped = tstamp;
}
@@ -783,6 +1140,7 @@ static int __perf_event_disable(void *in
*/
if (event->state >= PERF_EVENT_STATE_INACTIVE) {
update_context_time(ctx);
+ update_cgrp_time_from_event(event);
update_group_times(event);
if (event == event->group_leader)
group_sched_out(event, cpuctx, ctx);
@@ -851,6 +1209,41 @@ void perf_event_disable(struct perf_even
raw_spin_unlock_irq(&ctx->lock);
}
+static void perf_set_shadow_time(struct perf_event *event,
+ struct perf_event_context *ctx,
+ u64 tstamp)
+{
+ /*
+ * use the correct time source for the time snapshot
+ *
+ * We could get by without this by leveraging the
+ * fact that to get to this function, the caller
+ * has most likely already called update_context_time()
+ * and update_cgrp_time_xx() and thus both timestamp
+ * are identical (or very close). Given that tstamp is,
+ * already adjusted for cgroup, we could say that:
+ * tstamp - ctx->timestamp
+ * is equivalent to
+ * tstamp - cgrp->timestamp.
+ *
+ * Then, in perf_output_read(), the calculation would
+ * work with no changes because:
+ * - event is guaranteed scheduled in
+ * - no scheduled out in between
+ * - thus the timestamp would be the same
+ *
+ * But this is a bit hairy.
+ *
+ * So instead, we have an explicit cgroup call to remain
+ * within the time time source all along. We believe it
+ * is cleaner and simpler to understand.
+ */
+ if (is_cgroup_event(event))
+ perf_cgroup_set_shadow_time(event, tstamp);
+ else
+ event->shadow_ctx_time = tstamp - ctx->timestamp;
+}
+
static int
event_sched_in(struct perf_event *event,
struct perf_cpu_context *cpuctx,
@@ -876,7 +1269,7 @@ event_sched_in(struct perf_event *event,
event->tstamp_running += tstamp - event->tstamp_stopped;
- event->shadow_ctx_time = tstamp - ctx->timestamp;
+ perf_set_shadow_time(event, ctx, tstamp);
if (!is_software_event(event))
cpuctx->active_oncpu++;
@@ -992,12 +1385,13 @@ static void add_event_to_ctx(struct perf
list_add_event(event, ctx);
perf_group_attach(event);
- event->tstamp_enabled = tstamp;
event->tstamp_running = tstamp;
event->tstamp_stopped = tstamp;
+ event->tstamp_enabled = tstamp;
}
-static void perf_event_context_sched_in(struct perf_event_context *ctx);
+static void perf_event_context_sched_in(struct perf_event_context *ctx,
+ struct task_struct *tsk);
/*
* Cross CPU call to install and enable a performance event
@@ -1018,15 +1412,21 @@ static int __perf_install_in_context(vo
* which do context switches with IRQs enabled.
*/
if (ctx->task && !cpuctx->task_ctx)
- perf_event_context_sched_in(ctx);
+ perf_event_context_sched_in(ctx, ctx->task);
raw_spin_lock(&ctx->lock);
ctx->is_active = 1;
update_context_time(ctx);
+ /*
+ * update cgrp time only if current cgrp
+ * matches event->cgrp. Must be done before
+ * calling add_event_to_ctx()
+ */
+ update_cgrp_time_from_event(event);
add_event_to_ctx(event, ctx);
- if (!event_filter_match(event))
+ if (!event_filter_match(event, current))
goto unlock;
/*
@@ -1160,10 +1560,19 @@ static int __perf_event_enable(void *inf
if (event->state >= PERF_EVENT_STATE_INACTIVE)
goto unlock;
+
+ /*
+ * set current task's cgroup time reference point
+ */
+ perf_cgroup_set_timestamp(current, perf_clock());
+
__perf_event_mark_enabled(event, ctx);
- if (!event_filter_match(event))
+ if (!event_filter_match(event, current)) {
+ if (is_cgroup_event(event))
+ perf_cgroup_defer_enabled(event, current);
goto unlock;
+ }
/*
* If the event is in a group and isn't the group leader,
@@ -1292,6 +1701,7 @@ static void ctx_sched_out(struct perf_ev
if (likely(!ctx->nr_events))
goto out;
update_context_time(ctx);
+ update_cgrp_time_from_cpuctx(cpuctx);
if (!ctx->nr_active)
goto out;
@@ -1481,6 +1891,14 @@ void __perf_event_task_sched_out(struct
for_each_task_context_nr(ctxn)
perf_event_context_sched_out(task, ctxn, next);
+
+ /*
+ * if cgroup events exist on this CPU, then we need
+ * to check if we have to switch out PMU state.
+ * cgroup event are system-wide mode only
+ */
+ if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
+ perf_cgroup_sched_out(task);
}
static void task_ctx_sched_out(struct perf_event_context *ctx,
@@ -1509,16 +1927,21 @@ static void cpu_ctx_sched_out(struct per
static void
ctx_pinned_sched_in(struct perf_event_context *ctx,
- struct perf_cpu_context *cpuctx)
+ struct perf_cpu_context *cpuctx,
+ struct task_struct *task, int cgrp_sw)
{
struct perf_event *event;
list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
if (event->state <= PERF_EVENT_STATE_OFF)
continue;
- if (!event_filter_match(event))
+ if (!event_filter_match(event, task))
continue;
+ /* may need to reset tstamp_enabled */
+ if (is_cgroup_event(event))
+ perf_cgroup_mark_enabled(event, ctx);
+
if (group_can_go_on(event, cpuctx, 1))
group_sched_in(event, cpuctx, ctx);
@@ -1535,7 +1958,8 @@ ctx_pinned_sched_in(struct perf_event_co
static void
ctx_flexible_sched_in(struct perf_event_context *ctx,
- struct perf_cpu_context *cpuctx)
+ struct perf_cpu_context *cpuctx,
+ struct task_struct *task, int cgrp_sw)
{
struct perf_event *event;
int can_add_hw = 1;
@@ -1548,9 +1972,13 @@ ctx_flexible_sched_in(struct perf_event_
* Listen to the 'cpu' scheduling filter constraint
* of events:
*/
- if (!event_filter_match(event))
+ if (!event_filter_match(event, task))
continue;
+ /* may need to reset tstamp_enabled */
+ if (is_cgroup_event(event))
+ perf_cgroup_mark_enabled(event, ctx);
+
if (group_can_go_on(event, cpuctx, can_add_hw)) {
if (group_sched_in(event, cpuctx, ctx))
can_add_hw = 0;
@@ -1561,36 +1989,41 @@ ctx_flexible_sched_in(struct perf_event_
static void
ctx_sched_in(struct perf_event_context *ctx,
struct perf_cpu_context *cpuctx,
- enum event_type_t event_type)
+ enum event_type_t event_type,
+ struct task_struct *task, int cgrp_sw)
{
+ u64 now;
+
raw_spin_lock(&ctx->lock);
ctx->is_active = 1;
if (likely(!ctx->nr_events))
goto out;
- ctx->timestamp = perf_clock();
-
+ now = perf_clock();
+ ctx->timestamp = now;
+ perf_cgroup_set_timestamp(task, now);
/*
* First go through the list and put on any pinned groups
* in order to give them the best chance of going on.
*/
if (event_type & EVENT_PINNED)
- ctx_pinned_sched_in(ctx, cpuctx);
+ ctx_pinned_sched_in(ctx, cpuctx, task, cgrp_sw);
/* Then walk through the lower prio flexible groups */
if (event_type & EVENT_FLEXIBLE)
- ctx_flexible_sched_in(ctx, cpuctx);
+ ctx_flexible_sched_in(ctx, cpuctx, task, cgrp_sw);
out:
raw_spin_unlock(&ctx->lock);
}
static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
- enum event_type_t event_type)
+ enum event_type_t event_type,
+ struct task_struct *task, int cgrp_sw)
{
struct perf_event_context *ctx = &cpuctx->ctx;
- ctx_sched_in(ctx, cpuctx, event_type);
+ ctx_sched_in(ctx, cpuctx, event_type, task, cgrp_sw);
}
static void task_ctx_sched_in(struct perf_event_context *ctx,
@@ -1602,11 +2035,12 @@ static void task_ctx_sched_in(struct per
if (cpuctx->task_ctx == ctx)
return;
- ctx_sched_in(ctx, cpuctx, event_type);
+ ctx_sched_in(ctx, cpuctx, event_type, NULL, 0);
cpuctx->task_ctx = ctx;
}
-static void perf_event_context_sched_in(struct perf_event_context *ctx)
+static void perf_event_context_sched_in(struct perf_event_context *ctx,
+ struct task_struct *task)
{
struct perf_cpu_context *cpuctx;
@@ -1622,9 +2056,9 @@ static void perf_event_context_sched_in(
*/
cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
- ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
- cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
- ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
+ ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task, 0);
+ cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task, 0);
+ ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task, 0);
cpuctx->task_ctx = ctx;
@@ -1657,8 +2091,15 @@ void __perf_event_task_sched_in(struct t
if (likely(!ctx))
continue;
- perf_event_context_sched_in(ctx);
+ perf_event_context_sched_in(ctx, task);
}
+ /*
+ * if cgroup events exist on this CPU, then we need
+ * to check if we have to switch in PMU state.
+ * cgroup event are system-wide mode only
+ */
+ if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
+ perf_cgroup_sched_in(task);
}
#define MAX_INTERRUPTS (~0ULL)
@@ -1775,7 +2216,7 @@ static void perf_ctx_adjust_freq(struct
if (event->state != PERF_EVENT_STATE_ACTIVE)
continue;
- if (!event_filter_match(event))
+ if (!event_filter_match(event, current))
continue;
hwc = &event->hw;
@@ -1833,9 +2274,10 @@ static void perf_rotate_context(struct p
struct perf_event_context *ctx = NULL;
int rotate = 0, remove = 1;
- if (cpuctx->ctx.nr_events) {
+ ctx = &cpuctx->ctx;
+ if (ctx->nr_events) {
remove = 0;
- if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
+ if (ctx->nr_events != ctx->nr_active)
rotate = 1;
}
@@ -1862,7 +2304,7 @@ static void perf_rotate_context(struct p
if (ctx)
rotate_ctx(ctx);
- cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
+ cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, current, 0);
if (ctx)
task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
@@ -1941,7 +2383,7 @@ static void perf_event_enable_on_exec(st
raw_spin_unlock(&ctx->lock);
- perf_event_context_sched_in(ctx);
+ perf_event_context_sched_in(ctx, ctx->task);
out:
local_irq_restore(flags);
}
@@ -1968,6 +2410,7 @@ static void __perf_event_read(void *info
raw_spin_lock(&ctx->lock);
if (ctx->is_active)
update_context_time(ctx);
+ update_cgrp_time_from_event(event);
update_event_times(event);
if (event->state == PERF_EVENT_STATE_ACTIVE)
event->pmu->read(event);
@@ -1998,8 +2441,10 @@ static u64 perf_event_read(struct perf_e
* (e.g., thread is blocked), in that case
* we cannot update context time
*/
- if (ctx->is_active)
+ if (ctx->is_active) {
update_context_time(ctx);
+ update_cgrp_time_from_event(event);
+ }
update_event_times(event);
raw_spin_unlock_irqrestore(&ctx->lock, flags);
}
@@ -2384,7 +2829,7 @@ static void free_event(struct perf_event
if (!event->parent) {
if (event->attach_state & PERF_ATTACH_TASK)
- jump_label_dec(&perf_task_events);
+ jump_label_dec(&perf_sched_events);
if (event->attr.mmap || event->attr.mmap_data)
atomic_dec(&nr_mmap_events);
if (event->attr.comm)
@@ -2400,6 +2845,9 @@ static void free_event(struct perf_event
event->buffer = NULL;
}
+ if (is_cgroup_event(event))
+ perf_detach_cgroup(event);
+
if (event->destroy)
event->destroy(event);
@@ -3984,7 +4432,7 @@ static int perf_event_task_match(struct
if (event->state < PERF_EVENT_STATE_INACTIVE)
return 0;
- if (!event_filter_match(event))
+ if (!event_filter_match(event, current))
return 0;
if (event->attr.comm || event->attr.mmap ||
@@ -4121,7 +4569,7 @@ static int perf_event_comm_match(struct
if (event->state < PERF_EVENT_STATE_INACTIVE)
return 0;
- if (!event_filter_match(event))
+ if (!event_filter_match(event, current))
return 0;
if (event->attr.comm)
@@ -4269,7 +4717,7 @@ static int perf_event_mmap_match(struct
if (event->state < PERF_EVENT_STATE_INACTIVE)
return 0;
- if (!event_filter_match(event))
+ if (!event_filter_match(event, current))
return 0;
if ((!executable && event->attr.mmap_data) ||
@@ -5289,6 +5737,7 @@ static void task_clock_event_read(struct
if (!in_nmi()) {
update_context_time(event->ctx);
+ update_cgrp_time_from_event(event);
time = event->ctx->time;
} else {
u64 now = perf_clock();
@@ -5714,7 +6163,7 @@ perf_event_alloc(struct perf_event_attr
if (!event->parent) {
if (event->attach_state & PERF_ATTACH_TASK)
- jump_label_inc(&perf_task_events);
+ jump_label_inc(&perf_sched_events);
if (event->attr.mmap || event->attr.mmap_data)
atomic_inc(&nr_mmap_events);
if (event->attr.comm)
@@ -5889,7 +6338,7 @@ SYSCALL_DEFINE5(perf_event_open,
int err;
/* for future expandability... */
- if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
+ if (flags & ~PERF_FLAG_ALL)
return -EINVAL;
err = perf_copy_attr(attr_uptr, &attr);
@@ -5906,6 +6355,15 @@ SYSCALL_DEFINE5(perf_event_open,
return -EINVAL;
}
+ /*
+ * In cgroup mode, the pid argument is used to pass the fd
+ * opened to the cgroup directory in cgroupfs. The cpu argument
+ * designates the cpu on which to monitor threads from that
+ * cgroup.
+ */
+ if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
+ return -EINVAL;
+
event_fd = get_unused_fd_flags(O_RDWR);
if (event_fd < 0)
return event_fd;
@@ -5923,7 +6381,7 @@ SYSCALL_DEFINE5(perf_event_open,
group_leader = NULL;
}
- if (pid != -1) {
+ if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
task = find_lively_task_by_vpid(pid);
if (IS_ERR(task)) {
err = PTR_ERR(task);
@@ -5937,6 +6395,12 @@ SYSCALL_DEFINE5(perf_event_open,
goto err_task;
}
+ if (flags & PERF_FLAG_PID_CGROUP) {
+ err = perf_cgroup_connect(pid, event, &attr, group_leader);
+ if (err)
+ goto err_alloc;
+ }
+
/*
* Special case software events and allow them to be part of
* any hardware group.
@@ -6797,3 +7261,92 @@ static int __init perf_event_sysfs_init(
return ret;
}
device_initcall(perf_event_sysfs_init);
+
+#ifdef CONFIG_CGROUP_PERF
+static struct cgroup_subsys_state *perf_cgroup_create(
+ struct cgroup_subsys *ss, struct cgroup *cont)
+{
+ struct perf_cgroup *jc;
+ struct perf_cgroup_info *t;
+ int c;
+
+ jc = kmalloc(sizeof(*jc), GFP_KERNEL);
+ if (!jc)
+ return ERR_PTR(-ENOMEM);
+
+ memset(jc, 0, sizeof(*jc));
+
+ jc->info = alloc_percpu(struct perf_cgroup_info);
+ if (!jc->info) {
+ kfree(jc);
+ return ERR_PTR(-ENOMEM);
+ }
+
+ for_each_possible_cpu(c) {
+ t = per_cpu_ptr(jc->info, c);
+ t->time = 0;
+ t->timestamp = 0;
+ }
+ return &jc->css;
+}
+
+static void perf_cgroup_destroy(struct cgroup_subsys *ss,
+ struct cgroup *cont)
+{
+ struct perf_cgroup *jc;
+ jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
+ struct perf_cgroup, css);
+ free_percpu(jc->info);
+ kfree(jc);
+}
+
+static int __perf_cgroup_move(void *info)
+{
+ struct task_struct *task = info;
+ perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
+ return 0;
+}
+
+static void perf_cgroup_move(struct task_struct *task)
+{
+ task_function_call(task, __perf_cgroup_move, task);
+}
+
+static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
+ struct cgroup *old_cgrp, struct task_struct *task,
+ bool threadgroup)
+{
+ perf_cgroup_move(task);
+ if (threadgroup) {
+ struct task_struct *c;
+ rcu_read_lock();
+ list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
+ perf_cgroup_move(c);
+ }
+ rcu_read_unlock();
+ }
+}
+
+static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
+ struct cgroup *old_cgrp, struct task_struct *task)
+{
+ /*
+ * cgroup_exit() is called in the copy_process() failure path.
+ * Ignore this case since the task hasn't ran yet, this avoids
+ * trying to poke a half freed task state from generic code.
+ */
+ if (!(task->flags & PF_EXITING))
+ return;
+
+ perf_cgroup_move(task);
+}
+
+struct cgroup_subsys perf_subsys = {
+ .name = "perf_event",
+ .subsys_id = perf_subsys_id,
+ .create = perf_cgroup_create,
+ .destroy = perf_cgroup_destroy,
+ .exit = perf_cgroup_exit,
+ .attach = perf_cgroup_attach,
+};
+#endif /* CONFIG_CGROUP_PERF */
Peter,
I will try your changes and report back tomorrow.
Thanks.
On Mon, Feb 7, 2011 at 5:10 PM, Peter Zijlstra <[email protected]> wrote:
> Compile tested only, depends on the cgroup::exit patch
>
> ---
> Subject: perf: Add cgroup support
> From: Stephane Eranian <[email protected]>
> Date: Mon Feb 07 17:02:25 CET 2011
>
> This kernel patch adds the ability to filter monitoring based on
> container groups (cgroups). This is for use in per-cpu mode only.
>
> The cgroup to monitor is passed as a file descriptor in the pid
> argument to the syscall. The file descriptor must be opened to
> the cgroup name in the cgroup filesystem. For instance, if the
> cgroup name is foo and cgroupfs is mounted in /cgroup, then the
> file descriptor is opened to /cgroup/foo. Cgroup mode is
> activated by passing PERF_FLAG_PID_CGROUP in the flags argument
> to the syscall.
>
> For instance to measure in cgroup foo on CPU1 assuming
> cgroupfs is mounted under /cgroup:
>
> struct perf_event_attr attr;
> int cgroup_fd, fd;
>
> cgroup_fd = open("/cgroup/foo", O_RDONLY);
> fd = perf_event_open(&attr, cgroup_fd, 1, -1, PERF_FLAG_PID_CGROUP);
> close(cgroup_fd);
>
> Signed-off-by: Stephane Eranian <[email protected]>
> [ added perf_cgroup_{exit,attach} ]
> Signed-off-by: Peter Zijlstra <[email protected]>
> LKML-Reference: <new-submission>
> ---
> include/linux/cgroup.h | 1
> include/linux/cgroup_subsys.h | 4
> include/linux/perf_event.h | 33 +-
> init/Kconfig | 10
> kernel/cgroup.c | 23 +
> kernel/perf_event.c | 641 +++++++++++++++++++++++++++++++++++++++---
> 6 files changed, 665 insertions(+), 47 deletions(-)
>
> Index: linux-2.6/include/linux/cgroup.h
> ===================================================================
> --- linux-2.6.orig/include/linux/cgroup.h
> +++ linux-2.6/include/linux/cgroup.h
> @@ -627,6 +627,7 @@ bool css_is_ancestor(struct cgroup_subsy
> /* Get id and depth of css */
> unsigned short css_id(struct cgroup_subsys_state *css);
> unsigned short css_depth(struct cgroup_subsys_state *css);
> +struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id);
>
> #else /* !CONFIG_CGROUPS */
>
> Index: linux-2.6/include/linux/cgroup_subsys.h
> ===================================================================
> --- linux-2.6.orig/include/linux/cgroup_subsys.h
> +++ linux-2.6/include/linux/cgroup_subsys.h
> @@ -65,4 +65,8 @@ SUBSYS(net_cls)
> SUBSYS(blkio)
> #endif
>
> +#ifdef CONFIG_CGROUP_PERF
> +SUBSYS(perf)
> +#endif
> +
> /* */
> Index: linux-2.6/include/linux/perf_event.h
> ===================================================================
> --- linux-2.6.orig/include/linux/perf_event.h
> +++ linux-2.6/include/linux/perf_event.h
> @@ -464,6 +464,7 @@ enum perf_callchain_context {
>
> #define PERF_FLAG_FD_NO_GROUP (1U << 0)
> #define PERF_FLAG_FD_OUTPUT (1U << 1)
> +#define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */
>
> #ifdef __KERNEL__
> /*
> @@ -471,6 +472,7 @@ enum perf_callchain_context {
> */
>
> #ifdef CONFIG_PERF_EVENTS
> +# include <linux/cgroup.h>
> # include <asm/perf_event.h>
> # include <asm/local64.h>
> #endif
> @@ -716,6 +718,22 @@ struct swevent_hlist {
> #define PERF_ATTACH_GROUP 0x02
> #define PERF_ATTACH_TASK 0x04
>
> +#ifdef CONFIG_CGROUP_PERF
> +/*
> + * perf_cgroup_info keeps track of time_enabled for a cgroup.
> + * This is a per-cpu dynamically allocated data structure.
> + */
> +struct perf_cgroup_info {
> + u64 time;
> + u64 timestamp;
> +};
> +
> +struct perf_cgroup {
> + struct cgroup_subsys_state css;
> + struct perf_cgroup_info *info; /* timing info, one per cpu */
> +};
> +#endif
> +
> /**
> * struct perf_event - performance event kernel representation:
> */
> @@ -832,6 +850,11 @@ struct perf_event {
> struct event_filter *filter;
> #endif
>
> +#ifdef CONFIG_CGROUP_PERF
> + struct perf_cgroup *cgrp; /* cgroup event is attach to */
> + int cgrp_defer_enabled;
> +#endif
> +
> #endif /* CONFIG_PERF_EVENTS */
> };
>
> @@ -886,6 +909,7 @@ struct perf_event_context {
> u64 generation;
> int pin_count;
> struct rcu_head rcu_head;
> + int nr_cgroups; /* cgroup events present */
> };
>
> /*
> @@ -905,6 +929,9 @@ struct perf_cpu_context {
> struct list_head rotation_list;
> int jiffies_interval;
> struct pmu *active_pmu;
> +#ifdef CONFIG_CGROUP_PERF
> + struct perf_cgroup *cgrp;
> +#endif
> };
>
> struct perf_output_handle {
> @@ -1040,11 +1067,11 @@ perf_sw_event(u32 event_id, u64 nr, int
> __perf_sw_event(event_id, nr, nmi, regs, addr);
> }
>
> -extern atomic_t perf_task_events;
> +extern atomic_t perf_sched_events;
>
> static inline void perf_event_task_sched_in(struct task_struct *task)
> {
> - COND_STMT(&perf_task_events, __perf_event_task_sched_in(task));
> + COND_STMT(&perf_sched_events, __perf_event_task_sched_in(task));
> }
>
> static inline
> @@ -1052,7 +1079,7 @@ void perf_event_task_sched_out(struct ta
> {
> perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
>
> - COND_STMT(&perf_task_events, __perf_event_task_sched_out(task, next));
> + COND_STMT(&perf_sched_events, __perf_event_task_sched_out(task, next));
> }
>
> extern void perf_event_mmap(struct vm_area_struct *vma);
> Index: linux-2.6/init/Kconfig
> ===================================================================
> --- linux-2.6.orig/init/Kconfig
> +++ linux-2.6/init/Kconfig
> @@ -683,6 +683,16 @@ config CGROUP_MEM_RES_CTLR_SWAP_ENABLED
> select this option (if, for some reason, they need to disable it
> then noswapaccount does the trick).
>
> +config CGROUP_PERF
> + bool "Enable perf_event per-cpu per-container group (cgroup) monitoring"
> + depends on PERF_EVENTS && CGROUPS
> + help
> + This option extends the per-cpu mode to restrict monitoring to
> + threads which belong to the cgroup specificied and run on the
> + designated cpu.
> +
> + Say N if unsure.
> +
> menuconfig CGROUP_SCHED
> bool "Group CPU scheduler"
> depends on EXPERIMENTAL
> Index: linux-2.6/kernel/cgroup.c
> ===================================================================
> --- linux-2.6.orig/kernel/cgroup.c
> +++ linux-2.6/kernel/cgroup.c
> @@ -4822,6 +4822,29 @@ css_get_next(struct cgroup_subsys *ss, i
> return ret;
> }
>
> +/*
> + * get corresponding css from file open on cgroupfs directory
> + */
> +struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
> +{
> + struct cgroup *cgrp;
> + struct inode *inode;
> + struct cgroup_subsys_state *css;
> +
> + inode = f->f_dentry->d_inode;
> + /* check in cgroup filesystem dir */
> + if (inode->i_op != &cgroup_dir_inode_operations)
> + return ERR_PTR(-EBADF);
> +
> + if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
> + return ERR_PTR(-EINVAL);
> +
> + /* get cgroup */
> + cgrp = __d_cgrp(f->f_dentry);
> + css = cgrp->subsys[id];
> + return css ? css : ERR_PTR(-ENOENT);
> +}
> +
> #ifdef CONFIG_CGROUP_DEBUG
> static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
> struct cgroup *cont)
> Index: linux-2.6/kernel/perf_event.c
> ===================================================================
> --- linux-2.6.orig/kernel/perf_event.c
> +++ linux-2.6/kernel/perf_event.c
> @@ -111,13 +111,23 @@ static int cpu_function_call(int cpu, in
> return data.ret;
> }
>
> +#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
> + PERF_FLAG_FD_OUTPUT |\
> + PERF_FLAG_PID_CGROUP)
> +
> enum event_type_t {
> EVENT_FLEXIBLE = 0x1,
> EVENT_PINNED = 0x2,
> EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
> };
>
> -atomic_t perf_task_events __read_mostly;
> +/*
> + * perf_sched_events : >0 events exist
> + * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
> + */
> +atomic_t perf_sched_events __read_mostly;
> +static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
> +
> static atomic_t nr_mmap_events __read_mostly;
> static atomic_t nr_comm_events __read_mostly;
> static atomic_t nr_task_events __read_mostly;
> @@ -148,7 +158,11 @@ static void cpu_ctx_sched_out(struct per
> enum event_type_t event_type);
>
> static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
> - enum event_type_t event_type);
> + enum event_type_t event_type,
> + struct task_struct *task, int cgrp_sw);
> +
> +static void update_context_time(struct perf_event_context *ctx);
> +static u64 perf_event_time(struct perf_event *event);
>
> void __weak perf_event_print_debug(void) { }
>
> @@ -162,6 +176,315 @@ static inline u64 perf_clock(void)
> return local_clock();
> }
>
> +#ifdef CONFIG_CGROUP_PERF
> +
> +static inline struct perf_cgroup *
> +perf_cgroup_from_task(struct task_struct *task)
> +{
> + return container_of(task_subsys_state(task, perf_subsys_id),
> + struct perf_cgroup, css);
> +}
> +
> +static inline bool
> +perf_cgroup_match(struct perf_event *event, struct task_struct *task)
> +{
> + struct perf_cgroup *cgrp = NULL;
> + if (task)
> + cgrp = perf_cgroup_from_task(task);
> + return !event->cgrp || event->cgrp == cgrp;
> +}
> +
> +static inline void perf_get_cgroup(struct perf_event *event)
> +{
> + css_get(&event->cgrp->css);
> +}
> +
> +static inline void perf_put_cgroup(struct perf_event *event)
> +{
> + css_put(&event->cgrp->css);
> +}
> +
> +static inline void perf_detach_cgroup(struct perf_event *event)
> +{
> + perf_put_cgroup(event);
> + event->cgrp = NULL;
> +}
> +
> +static inline int is_cgroup_event(struct perf_event *event)
> +{
> + return event->cgrp != NULL;
> +}
> +
> +static inline u64 perf_cgroup_event_time(struct perf_event *event)
> +{
> + struct perf_cgroup_info *t;
> +
> + t = per_cpu_ptr(event->cgrp->info, event->cpu);
> + return t->time;
> +}
> +
> +static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
> +{
> + struct perf_cgroup_info *info;
> + u64 now;
> +
> + now = perf_clock();
> +
> + info = this_cpu_ptr(cgrp->info);
> +
> + info->time += now - info->timestamp;
> + info->timestamp = now;
> +}
> +
> +static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
> +{
> + struct perf_cgroup *cgrp_out = cpuctx->cgrp;
> + if (cgrp_out)
> + __update_cgrp_time(cgrp_out);
> +}
> +
> +static inline void update_cgrp_time_from_event(struct perf_event *event)
> +{
> + struct perf_cgroup *cgrp = perf_cgroup_from_task(current);
> + /*
> + * do not update time when cgroup is not active
> + */
> + if (!event->cgrp || cgrp != event->cgrp)
> + return;
> +
> + __update_cgrp_time(event->cgrp);
> +}
> +
> +static inline void
> +perf_cgroup_set_timestamp(struct task_struct *task, u64 now)
> +{
> + struct perf_cgroup *cgrp;
> + struct perf_cgroup_info *info;
> +
> + if (!task)
> + return;
> +
> + cgrp = perf_cgroup_from_task(task);
> + info = per_cpu_ptr(cgrp->info, smp_processor_id());
> + info->timestamp = now;
> +}
> +
> +#define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
> +#define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
> +
> +/*
> + * reschedule events based on the cgroup constraint of task.
> + *
> + * mode SWOUT : schedule out everything
> + * mode SWIN : schedule in based on cgroup for next
> + */
> +void perf_cgroup_switch(struct task_struct *task, int mode)
> +{
> + struct perf_cpu_context *cpuctx;
> + struct pmu *pmu;
> + unsigned long flags;
> +
> + /*
> + * disable interrupts to avoid geting nr_cgroup
> + * changes via __perf_event_disable(). Also
> + * avoids preemption.
> + */
> + local_irq_save(flags);
> +
> + /*
> + * we reschedule only in the presence of cgroup
> + * constrained events.
> + */
> + rcu_read_lock();
> +
> + list_for_each_entry_rcu(pmu, &pmus, entry) {
> +
> + cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
> +
> + perf_pmu_disable(cpuctx->ctx.pmu);
> +
> + /*
> + * perf_cgroup_events says at least one
> + * context on this CPU has cgroup events.
> + *
> + * ctx->nr_cgroups reports the number of cgroup
> + * events for a context.
> + */
> + if (cpuctx->ctx.nr_cgroups > 0) {
> +
> + if (mode & PERF_CGROUP_SWOUT)
> + cpu_ctx_sched_out(cpuctx, EVENT_ALL);
> +
> + if (mode & PERF_CGROUP_SWIN) {
> + cpu_ctx_sched_in(cpuctx, EVENT_ALL, task, 1);
> + cpuctx->cgrp = perf_cgroup_from_task(task);
> + }
> + }
> +
> + perf_pmu_enable(cpuctx->ctx.pmu);
> + }
> +
> + rcu_read_unlock();
> +
> + local_irq_restore(flags);
> +}
> +
> +static inline void perf_cgroup_sched_out(struct task_struct *task)
> +{
> + perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
> +}
> +
> +static inline void perf_cgroup_sched_in(struct task_struct *task)
> +{
> + perf_cgroup_switch(task, PERF_CGROUP_SWIN);
> +}
> +
> +static inline int perf_cgroup_connect(int fd, struct perf_event *event,
> + struct perf_event_attr *attr,
> + struct perf_event *group_leader)
> +{
> + struct perf_cgroup *cgrp;
> + struct cgroup_subsys_state *css;
> + struct file *file;
> + int ret = 0, fput_needed;
> +
> + file = fget_light(fd, &fput_needed);
> + if (!file)
> + return -EBADF;
> +
> + css = cgroup_css_from_dir(file, perf_subsys_id);
> + if (IS_ERR(css))
> + return PTR_ERR(css);
> +
> + cgrp = container_of(css, struct perf_cgroup, css);
> + event->cgrp = cgrp;
> +
> + /*
> + * all events in a group must monitor
> + * the same cgroup because a task belongs
> + * to only one perf cgroup at a time
> + */
> + if (group_leader && group_leader->cgrp != cgrp) {
> + perf_detach_cgroup(event);
> + ret = -EINVAL;
> + } else {
> + /* must be done before we fput() the file */
> + perf_get_cgroup(event);
> + }
> + fput_light(file, fput_needed);
> + return ret;
> +}
> +
> +static inline void
> +perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
> +{
> + struct perf_cgroup_info *t;
> + t = per_cpu_ptr(event->cgrp->info, event->cpu);
> + event->shadow_ctx_time = now - t->timestamp;
> +}
> +
> +static inline void
> +perf_cgroup_defer_enabled(struct perf_event *event, struct task_struct *task)
> +{
> + /*
> + * when the current task's perf cgroup does not match
> + * the event's, we need to remember to call the
> + * perf_mark_enable() function the first time a task with
> + * a matching perf cgroup is scheduled in.
> + */
> + if (is_cgroup_event(event) && !perf_cgroup_match(event, task))
> + event->cgrp_defer_enabled = 1;
> +}
> +
> +static inline void
> +perf_cgroup_mark_enabled(struct perf_event *event,
> + struct perf_event_context *ctx)
> +{
> + struct perf_event *sub;
> + u64 tstamp = perf_event_time(event);
> +
> + if (!event->cgrp_defer_enabled)
> + return;
> +
> + event->cgrp_defer_enabled = 0;
> +
> + event->tstamp_enabled = tstamp - event->total_time_enabled;
> + list_for_each_entry(sub, &event->sibling_list, group_entry) {
> + if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
> + sub->tstamp_enabled = tstamp - sub->total_time_enabled;
> + sub->cgrp_defer_enabled = 0;
> + }
> + }
> +}
> +#else /* !CONFIG_CGROUP_PERF */
> +
> +static inline bool
> +perf_cgroup_match(struct perf_event *event, struct task_struct *task)
> +{
> + return true;
> +}
> +
> +static inline void perf_detach_cgroup(struct perf_event *event)
> +{}
> +
> +static inline int is_cgroup_event(struct perf_event *event)
> +{
> + return 0;
> +}
> +
> +static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
> +{
> + return 0;
> +}
> +
> +static inline void update_cgrp_time_from_event(struct perf_event *event)
> +{}
> +
> +static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
> +{}
> +
> +static inline void perf_cgroup_sched_out(struct task_struct *task)
> +{
> +}
> +
> +static inline void perf_cgroup_sched_in(struct task_struct *task)
> +{
> +}
> +
> +static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
> + struct perf_event_attr *attr,
> + struct perf_event *group_leader)
> +{
> + return -EINVAL;
> +}
> +
> +static inline void
> +perf_cgroup_set_timestamp(struct task_struct *task, u64 now)
> +{}
> +
> +void
> +perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
> +{}
> +
> +static inline void
> +perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
> +{}
> +
> +static inline u64 perf_cgroup_event_time(struct perf_event *event)
> +{
> + return 0;
> +}
> +
> +static inline void
> +perf_cgroup_defer_enabled(struct perf_event *event, struct task_struct *task)
> +{}
> +
> +static inline void
> +perf_cgroup_mark_enabled(struct perf_event *event,
> + struct perf_event_context *ctx)
> +{}
> +#endif
> +
> void perf_pmu_disable(struct pmu *pmu)
> {
> int *count = this_cpu_ptr(pmu->pmu_disable_count);
> @@ -343,6 +666,10 @@ static void update_context_time(struct p
> static u64 perf_event_time(struct perf_event *event)
> {
> struct perf_event_context *ctx = event->ctx;
> +
> + if (is_cgroup_event(event))
> + return perf_cgroup_event_time(event);
> +
> return ctx ? ctx->time : 0;
> }
>
> @@ -357,9 +684,20 @@ static void update_event_times(struct pe
> if (event->state < PERF_EVENT_STATE_INACTIVE ||
> event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
> return;
> -
> - if (ctx->is_active)
> + /*
> + * in cgroup mode, time_enabled represents
> + * the time the event was enabled AND active
> + * tasks were in the monitored cgroup. This is
> + * independent of the activity of the context as
> + * there may be a mix of cgroup and non-cgroup events.
> + *
> + * That is why we treat cgroup events differently
> + * here.
> + */
> + if (is_cgroup_event(event))
> run_end = perf_event_time(event);
> + else if (ctx->is_active)
> + run_end = ctx->time;
> else
> run_end = event->tstamp_stopped;
>
> @@ -371,6 +709,7 @@ static void update_event_times(struct pe
> run_end = perf_event_time(event);
>
> event->total_time_running = run_end - event->tstamp_running;
> +
> }
>
> /*
> @@ -419,6 +758,17 @@ list_add_event(struct perf_event *event,
> list_add_tail(&event->group_entry, list);
> }
>
> + if (is_cgroup_event(event)) {
> + ctx->nr_cgroups++;
> + /*
> + * one more event:
> + * - that has cgroup constraint on event->cpu
> + * - that may need work on context switch
> + */
> + atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
> + jump_label_inc(&perf_sched_events);
> + }
> +
> list_add_rcu(&event->event_entry, &ctx->event_list);
> if (!ctx->nr_events)
> perf_pmu_rotate_start(ctx->pmu);
> @@ -545,6 +895,12 @@ list_del_event(struct perf_event *event,
>
> event->attach_state &= ~PERF_ATTACH_CONTEXT;
>
> + if (is_cgroup_event(event)) {
> + ctx->nr_cgroups--;
> + atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
> + jump_label_dec(&perf_sched_events);
> + }
> +
> ctx->nr_events--;
> if (event->attr.inherit_stat)
> ctx->nr_stat--;
> @@ -614,9 +970,10 @@ static void perf_group_detach(struct per
> }
>
> static inline int
> -event_filter_match(struct perf_event *event)
> +event_filter_match(struct perf_event *event, struct task_struct *task)
> {
> - return event->cpu == -1 || event->cpu == smp_processor_id();
> + return (event->cpu == -1 || event->cpu == smp_processor_id())
> + && perf_cgroup_match(event, task);
> }
>
> static void
> @@ -633,8 +990,8 @@ event_sched_out(struct perf_event *event
> * via read() for time_enabled, time_running:
> */
> if (event->state == PERF_EVENT_STATE_INACTIVE
> - && !event_filter_match(event)) {
> - delta = ctx->time - event->tstamp_stopped;
> + && !event_filter_match(event, current)) {
> + delta = tstamp - event->tstamp_stopped;
> event->tstamp_running += delta;
> event->tstamp_stopped = tstamp;
> }
> @@ -783,6 +1140,7 @@ static int __perf_event_disable(void *in
> */
> if (event->state >= PERF_EVENT_STATE_INACTIVE) {
> update_context_time(ctx);
> + update_cgrp_time_from_event(event);
> update_group_times(event);
> if (event == event->group_leader)
> group_sched_out(event, cpuctx, ctx);
> @@ -851,6 +1209,41 @@ void perf_event_disable(struct perf_even
> raw_spin_unlock_irq(&ctx->lock);
> }
>
> +static void perf_set_shadow_time(struct perf_event *event,
> + struct perf_event_context *ctx,
> + u64 tstamp)
> +{
> + /*
> + * use the correct time source for the time snapshot
> + *
> + * We could get by without this by leveraging the
> + * fact that to get to this function, the caller
> + * has most likely already called update_context_time()
> + * and update_cgrp_time_xx() and thus both timestamp
> + * are identical (or very close). Given that tstamp is,
> + * already adjusted for cgroup, we could say that:
> + * tstamp - ctx->timestamp
> + * is equivalent to
> + * tstamp - cgrp->timestamp.
> + *
> + * Then, in perf_output_read(), the calculation would
> + * work with no changes because:
> + * - event is guaranteed scheduled in
> + * - no scheduled out in between
> + * - thus the timestamp would be the same
> + *
> + * But this is a bit hairy.
> + *
> + * So instead, we have an explicit cgroup call to remain
> + * within the time time source all along. We believe it
> + * is cleaner and simpler to understand.
> + */
> + if (is_cgroup_event(event))
> + perf_cgroup_set_shadow_time(event, tstamp);
> + else
> + event->shadow_ctx_time = tstamp - ctx->timestamp;
> +}
> +
> static int
> event_sched_in(struct perf_event *event,
> struct perf_cpu_context *cpuctx,
> @@ -876,7 +1269,7 @@ event_sched_in(struct perf_event *event,
>
> event->tstamp_running += tstamp - event->tstamp_stopped;
>
> - event->shadow_ctx_time = tstamp - ctx->timestamp;
> + perf_set_shadow_time(event, ctx, tstamp);
>
> if (!is_software_event(event))
> cpuctx->active_oncpu++;
> @@ -992,12 +1385,13 @@ static void add_event_to_ctx(struct perf
>
> list_add_event(event, ctx);
> perf_group_attach(event);
> - event->tstamp_enabled = tstamp;
> event->tstamp_running = tstamp;
> event->tstamp_stopped = tstamp;
> + event->tstamp_enabled = tstamp;
> }
>
> -static void perf_event_context_sched_in(struct perf_event_context *ctx);
> +static void perf_event_context_sched_in(struct perf_event_context *ctx,
> + struct task_struct *tsk);
>
> /*
> * Cross CPU call to install and enable a performance event
> @@ -1018,15 +1412,21 @@ static int __perf_install_in_context(vo
> * which do context switches with IRQs enabled.
> */
> if (ctx->task && !cpuctx->task_ctx)
> - perf_event_context_sched_in(ctx);
> + perf_event_context_sched_in(ctx, ctx->task);
>
> raw_spin_lock(&ctx->lock);
> ctx->is_active = 1;
> update_context_time(ctx);
> + /*
> + * update cgrp time only if current cgrp
> + * matches event->cgrp. Must be done before
> + * calling add_event_to_ctx()
> + */
> + update_cgrp_time_from_event(event);
>
> add_event_to_ctx(event, ctx);
>
> - if (!event_filter_match(event))
> + if (!event_filter_match(event, current))
> goto unlock;
>
> /*
> @@ -1160,10 +1560,19 @@ static int __perf_event_enable(void *inf
>
> if (event->state >= PERF_EVENT_STATE_INACTIVE)
> goto unlock;
> +
> + /*
> + * set current task's cgroup time reference point
> + */
> + perf_cgroup_set_timestamp(current, perf_clock());
> +
> __perf_event_mark_enabled(event, ctx);
>
> - if (!event_filter_match(event))
> + if (!event_filter_match(event, current)) {
> + if (is_cgroup_event(event))
> + perf_cgroup_defer_enabled(event, current);
> goto unlock;
> + }
>
> /*
> * If the event is in a group and isn't the group leader,
> @@ -1292,6 +1701,7 @@ static void ctx_sched_out(struct perf_ev
> if (likely(!ctx->nr_events))
> goto out;
> update_context_time(ctx);
> + update_cgrp_time_from_cpuctx(cpuctx);
>
> if (!ctx->nr_active)
> goto out;
> @@ -1481,6 +1891,14 @@ void __perf_event_task_sched_out(struct
>
> for_each_task_context_nr(ctxn)
> perf_event_context_sched_out(task, ctxn, next);
> +
> + /*
> + * if cgroup events exist on this CPU, then we need
> + * to check if we have to switch out PMU state.
> + * cgroup event are system-wide mode only
> + */
> + if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
> + perf_cgroup_sched_out(task);
> }
>
> static void task_ctx_sched_out(struct perf_event_context *ctx,
> @@ -1509,16 +1927,21 @@ static void cpu_ctx_sched_out(struct per
>
> static void
> ctx_pinned_sched_in(struct perf_event_context *ctx,
> - struct perf_cpu_context *cpuctx)
> + struct perf_cpu_context *cpuctx,
> + struct task_struct *task, int cgrp_sw)
> {
> struct perf_event *event;
>
> list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
> if (event->state <= PERF_EVENT_STATE_OFF)
> continue;
> - if (!event_filter_match(event))
> + if (!event_filter_match(event, task))
> continue;
>
> + /* may need to reset tstamp_enabled */
> + if (is_cgroup_event(event))
> + perf_cgroup_mark_enabled(event, ctx);
> +
> if (group_can_go_on(event, cpuctx, 1))
> group_sched_in(event, cpuctx, ctx);
>
> @@ -1535,7 +1958,8 @@ ctx_pinned_sched_in(struct perf_event_co
>
> static void
> ctx_flexible_sched_in(struct perf_event_context *ctx,
> - struct perf_cpu_context *cpuctx)
> + struct perf_cpu_context *cpuctx,
> + struct task_struct *task, int cgrp_sw)
> {
> struct perf_event *event;
> int can_add_hw = 1;
> @@ -1548,9 +1972,13 @@ ctx_flexible_sched_in(struct perf_event_
> * Listen to the 'cpu' scheduling filter constraint
> * of events:
> */
> - if (!event_filter_match(event))
> + if (!event_filter_match(event, task))
> continue;
>
> + /* may need to reset tstamp_enabled */
> + if (is_cgroup_event(event))
> + perf_cgroup_mark_enabled(event, ctx);
> +
> if (group_can_go_on(event, cpuctx, can_add_hw)) {
> if (group_sched_in(event, cpuctx, ctx))
> can_add_hw = 0;
> @@ -1561,36 +1989,41 @@ ctx_flexible_sched_in(struct perf_event_
> static void
> ctx_sched_in(struct perf_event_context *ctx,
> struct perf_cpu_context *cpuctx,
> - enum event_type_t event_type)
> + enum event_type_t event_type,
> + struct task_struct *task, int cgrp_sw)
> {
> + u64 now;
> +
> raw_spin_lock(&ctx->lock);
> ctx->is_active = 1;
> if (likely(!ctx->nr_events))
> goto out;
>
> - ctx->timestamp = perf_clock();
> -
> + now = perf_clock();
> + ctx->timestamp = now;
> + perf_cgroup_set_timestamp(task, now);
> /*
> * First go through the list and put on any pinned groups
> * in order to give them the best chance of going on.
> */
> if (event_type & EVENT_PINNED)
> - ctx_pinned_sched_in(ctx, cpuctx);
> + ctx_pinned_sched_in(ctx, cpuctx, task, cgrp_sw);
>
> /* Then walk through the lower prio flexible groups */
> if (event_type & EVENT_FLEXIBLE)
> - ctx_flexible_sched_in(ctx, cpuctx);
> + ctx_flexible_sched_in(ctx, cpuctx, task, cgrp_sw);
>
> out:
> raw_spin_unlock(&ctx->lock);
> }
>
> static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
> - enum event_type_t event_type)
> + enum event_type_t event_type,
> + struct task_struct *task, int cgrp_sw)
> {
> struct perf_event_context *ctx = &cpuctx->ctx;
>
> - ctx_sched_in(ctx, cpuctx, event_type);
> + ctx_sched_in(ctx, cpuctx, event_type, task, cgrp_sw);
> }
>
> static void task_ctx_sched_in(struct perf_event_context *ctx,
> @@ -1602,11 +2035,12 @@ static void task_ctx_sched_in(struct per
> if (cpuctx->task_ctx == ctx)
> return;
>
> - ctx_sched_in(ctx, cpuctx, event_type);
> + ctx_sched_in(ctx, cpuctx, event_type, NULL, 0);
> cpuctx->task_ctx = ctx;
> }
>
> -static void perf_event_context_sched_in(struct perf_event_context *ctx)
> +static void perf_event_context_sched_in(struct perf_event_context *ctx,
> + struct task_struct *task)
> {
> struct perf_cpu_context *cpuctx;
>
> @@ -1622,9 +2056,9 @@ static void perf_event_context_sched_in(
> */
> cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
>
> - ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
> - cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
> - ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
> + ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task, 0);
> + cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task, 0);
> + ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task, 0);
>
> cpuctx->task_ctx = ctx;
>
> @@ -1657,8 +2091,15 @@ void __perf_event_task_sched_in(struct t
> if (likely(!ctx))
> continue;
>
> - perf_event_context_sched_in(ctx);
> + perf_event_context_sched_in(ctx, task);
> }
> + /*
> + * if cgroup events exist on this CPU, then we need
> + * to check if we have to switch in PMU state.
> + * cgroup event are system-wide mode only
> + */
> + if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
> + perf_cgroup_sched_in(task);
> }
>
> #define MAX_INTERRUPTS (~0ULL)
> @@ -1775,7 +2216,7 @@ static void perf_ctx_adjust_freq(struct
> if (event->state != PERF_EVENT_STATE_ACTIVE)
> continue;
>
> - if (!event_filter_match(event))
> + if (!event_filter_match(event, current))
> continue;
>
> hwc = &event->hw;
> @@ -1833,9 +2274,10 @@ static void perf_rotate_context(struct p
> struct perf_event_context *ctx = NULL;
> int rotate = 0, remove = 1;
>
> - if (cpuctx->ctx.nr_events) {
> + ctx = &cpuctx->ctx;
> + if (ctx->nr_events) {
> remove = 0;
> - if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
> + if (ctx->nr_events != ctx->nr_active)
> rotate = 1;
> }
>
> @@ -1862,7 +2304,7 @@ static void perf_rotate_context(struct p
> if (ctx)
> rotate_ctx(ctx);
>
> - cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
> + cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, current, 0);
> if (ctx)
> task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
>
> @@ -1941,7 +2383,7 @@ static void perf_event_enable_on_exec(st
>
> raw_spin_unlock(&ctx->lock);
>
> - perf_event_context_sched_in(ctx);
> + perf_event_context_sched_in(ctx, ctx->task);
> out:
> local_irq_restore(flags);
> }
> @@ -1968,6 +2410,7 @@ static void __perf_event_read(void *info
> raw_spin_lock(&ctx->lock);
> if (ctx->is_active)
> update_context_time(ctx);
> + update_cgrp_time_from_event(event);
> update_event_times(event);
> if (event->state == PERF_EVENT_STATE_ACTIVE)
> event->pmu->read(event);
> @@ -1998,8 +2441,10 @@ static u64 perf_event_read(struct perf_e
> * (e.g., thread is blocked), in that case
> * we cannot update context time
> */
> - if (ctx->is_active)
> + if (ctx->is_active) {
> update_context_time(ctx);
> + update_cgrp_time_from_event(event);
> + }
> update_event_times(event);
> raw_spin_unlock_irqrestore(&ctx->lock, flags);
> }
> @@ -2384,7 +2829,7 @@ static void free_event(struct perf_event
>
> if (!event->parent) {
> if (event->attach_state & PERF_ATTACH_TASK)
> - jump_label_dec(&perf_task_events);
> + jump_label_dec(&perf_sched_events);
> if (event->attr.mmap || event->attr.mmap_data)
> atomic_dec(&nr_mmap_events);
> if (event->attr.comm)
> @@ -2400,6 +2845,9 @@ static void free_event(struct perf_event
> event->buffer = NULL;
> }
>
> + if (is_cgroup_event(event))
> + perf_detach_cgroup(event);
> +
> if (event->destroy)
> event->destroy(event);
>
> @@ -3984,7 +4432,7 @@ static int perf_event_task_match(struct
> if (event->state < PERF_EVENT_STATE_INACTIVE)
> return 0;
>
> - if (!event_filter_match(event))
> + if (!event_filter_match(event, current))
> return 0;
>
> if (event->attr.comm || event->attr.mmap ||
> @@ -4121,7 +4569,7 @@ static int perf_event_comm_match(struct
> if (event->state < PERF_EVENT_STATE_INACTIVE)
> return 0;
>
> - if (!event_filter_match(event))
> + if (!event_filter_match(event, current))
> return 0;
>
> if (event->attr.comm)
> @@ -4269,7 +4717,7 @@ static int perf_event_mmap_match(struct
> if (event->state < PERF_EVENT_STATE_INACTIVE)
> return 0;
>
> - if (!event_filter_match(event))
> + if (!event_filter_match(event, current))
> return 0;
>
> if ((!executable && event->attr.mmap_data) ||
> @@ -5289,6 +5737,7 @@ static void task_clock_event_read(struct
>
> if (!in_nmi()) {
> update_context_time(event->ctx);
> + update_cgrp_time_from_event(event);
> time = event->ctx->time;
> } else {
> u64 now = perf_clock();
> @@ -5714,7 +6163,7 @@ perf_event_alloc(struct perf_event_attr
>
> if (!event->parent) {
> if (event->attach_state & PERF_ATTACH_TASK)
> - jump_label_inc(&perf_task_events);
> + jump_label_inc(&perf_sched_events);
> if (event->attr.mmap || event->attr.mmap_data)
> atomic_inc(&nr_mmap_events);
> if (event->attr.comm)
> @@ -5889,7 +6338,7 @@ SYSCALL_DEFINE5(perf_event_open,
> int err;
>
> /* for future expandability... */
> - if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
> + if (flags & ~PERF_FLAG_ALL)
> return -EINVAL;
>
> err = perf_copy_attr(attr_uptr, &attr);
> @@ -5906,6 +6355,15 @@ SYSCALL_DEFINE5(perf_event_open,
> return -EINVAL;
> }
>
> + /*
> + * In cgroup mode, the pid argument is used to pass the fd
> + * opened to the cgroup directory in cgroupfs. The cpu argument
> + * designates the cpu on which to monitor threads from that
> + * cgroup.
> + */
> + if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
> + return -EINVAL;
> +
> event_fd = get_unused_fd_flags(O_RDWR);
> if (event_fd < 0)
> return event_fd;
> @@ -5923,7 +6381,7 @@ SYSCALL_DEFINE5(perf_event_open,
> group_leader = NULL;
> }
>
> - if (pid != -1) {
> + if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
> task = find_lively_task_by_vpid(pid);
> if (IS_ERR(task)) {
> err = PTR_ERR(task);
> @@ -5937,6 +6395,12 @@ SYSCALL_DEFINE5(perf_event_open,
> goto err_task;
> }
>
> + if (flags & PERF_FLAG_PID_CGROUP) {
> + err = perf_cgroup_connect(pid, event, &attr, group_leader);
> + if (err)
> + goto err_alloc;
> + }
> +
> /*
> * Special case software events and allow them to be part of
> * any hardware group.
> @@ -6797,3 +7261,92 @@ static int __init perf_event_sysfs_init(
> return ret;
> }
> device_initcall(perf_event_sysfs_init);
> +
> +#ifdef CONFIG_CGROUP_PERF
> +static struct cgroup_subsys_state *perf_cgroup_create(
> + struct cgroup_subsys *ss, struct cgroup *cont)
> +{
> + struct perf_cgroup *jc;
> + struct perf_cgroup_info *t;
> + int c;
> +
> + jc = kmalloc(sizeof(*jc), GFP_KERNEL);
> + if (!jc)
> + return ERR_PTR(-ENOMEM);
> +
> + memset(jc, 0, sizeof(*jc));
> +
> + jc->info = alloc_percpu(struct perf_cgroup_info);
> + if (!jc->info) {
> + kfree(jc);
> + return ERR_PTR(-ENOMEM);
> + }
> +
> + for_each_possible_cpu(c) {
> + t = per_cpu_ptr(jc->info, c);
> + t->time = 0;
> + t->timestamp = 0;
> + }
> + return &jc->css;
> +}
> +
> +static void perf_cgroup_destroy(struct cgroup_subsys *ss,
> + struct cgroup *cont)
> +{
> + struct perf_cgroup *jc;
> + jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
> + struct perf_cgroup, css);
> + free_percpu(jc->info);
> + kfree(jc);
> +}
> +
> +static int __perf_cgroup_move(void *info)
> +{
> + struct task_struct *task = info;
> + perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
> + return 0;
> +}
> +
> +static void perf_cgroup_move(struct task_struct *task)
> +{
> + task_function_call(task, __perf_cgroup_move, task);
> +}
> +
> +static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
> + struct cgroup *old_cgrp, struct task_struct *task,
> + bool threadgroup)
> +{
> + perf_cgroup_move(task);
> + if (threadgroup) {
> + struct task_struct *c;
> + rcu_read_lock();
> + list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
> + perf_cgroup_move(c);
> + }
> + rcu_read_unlock();
> + }
> +}
> +
> +static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
> + struct cgroup *old_cgrp, struct task_struct *task)
> +{
> + /*
> + * cgroup_exit() is called in the copy_process() failure path.
> + * Ignore this case since the task hasn't ran yet, this avoids
> + * trying to poke a half freed task state from generic code.
> + */
> + if (!(task->flags & PF_EXITING))
> + return;
> +
> + perf_cgroup_move(task);
> +}
> +
> +struct cgroup_subsys perf_subsys = {
> + .name = "perf_event",
> + .subsys_id = perf_subsys_id,
> + .create = perf_cgroup_create,
> + .destroy = perf_cgroup_destroy,
> + .exit = perf_cgroup_exit,
> + .attach = perf_cgroup_attach,
> +};
> +#endif /* CONFIG_CGROUP_PERF */
>
>
>
Peter,
See comments below.
On Mon, Feb 7, 2011 at 5:10 PM, Peter Zijlstra <[email protected]> wrote:
> Compile tested only, depends on the cgroup::exit patch
>
> --- linux-2.6.orig/include/linux/perf_event.h
> +++ linux-2.6/include/linux/perf_event.h
> @@ -905,6 +929,9 @@ struct perf_cpu_context {
> struct list_head rotation_list;
> int jiffies_interval;
> struct pmu *active_pmu;
> +#ifdef CONFIG_CGROUP_PERF
> + struct perf_cgroup *cgrp;
> +#endif
> };
>
I don't quite understand the motivation for adding cgrp to cpuctx.
> --- linux-2.6.orig/kernel/perf_event.c
> +++ linux-2.6/kernel/perf_event.c
> +static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
> +{
> + struct perf_cgroup *cgrp_out = cpuctx->cgrp;
> + if (cgrp_out)
> + __update_cgrp_time(cgrp_out);
> +}
> +
What's the benefit of this form compared to the original from_task() version?
> +void perf_cgroup_switch(struct task_struct *task, int mode)
> +{
> + struct perf_cpu_context *cpuctx;
> + struct pmu *pmu;
> + unsigned long flags;
> +
> + /*
> + * disable interrupts to avoid geting nr_cgroup
> + * changes via __perf_event_disable(). Also
> + * avoids preemption.
> + */
> + local_irq_save(flags);
> +
> + /*
> + * we reschedule only in the presence of cgroup
> + * constrained events.
> + */
> + rcu_read_lock();
> +
> + list_for_each_entry_rcu(pmu, &pmus, entry) {
> +
> + cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
> +
> + perf_pmu_disable(cpuctx->ctx.pmu);
> +
> + /*
> + * perf_cgroup_events says at least one
> + * context on this CPU has cgroup events.
> + *
> + * ctx->nr_cgroups reports the number of cgroup
> + * events for a context.
> + */
> + if (cpuctx->ctx.nr_cgroups > 0) {
> +
> + if (mode & PERF_CGROUP_SWOUT)
> + cpu_ctx_sched_out(cpuctx, EVENT_ALL);
> +
> + if (mode & PERF_CGROUP_SWIN) {
> + cpu_ctx_sched_in(cpuctx, EVENT_ALL, task, 1);
> + cpuctx->cgrp = perf_cgroup_from_task(task);
> + }
> + }
I think there is a risk on cpuctx->cgrp pointing to stale cgrp information.
Shouldn't we also set cpuctx->cgrp = NULL on SWOUT?
> +static int __perf_cgroup_move(void *info)
> +{
> + struct task_struct *task = info;
> + perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
> + return 0;
> +}
> +
> +static void perf_cgroup_move(struct task_struct *task)
> +{
> + task_function_call(task, __perf_cgroup_move, task);
> +}
> +
> +static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
> + struct cgroup *old_cgrp, struct task_struct *task,
> + bool threadgroup)
> +{
> + perf_cgroup_move(task);
> + if (threadgroup) {
> + struct task_struct *c;
> + rcu_read_lock();
> + list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
> + perf_cgroup_move(c);
> + }
> + rcu_read_unlock();
> + }
> +}
> +
I suspect my original patch was not necessarily handling the attach completely
when you move an existing task into a cgroup which was already monitored.
I think you may have had to wait until a ctxsw. Looks like this callback handles
this better.
Let me make sure I understand the threadgroup iteration, though. I suspect
this handles the situation where a multi-threaded app is moved into a cgroup
while there is already cgroup monitoring active. In that case and if we do not
want to wait until there is at least one ctxsw on all CPUs, then we have to
check if the other threads are not already running on the other CPUs.If so,
we need to do a cgroup switch on those CPUs. Otherwise, we have nothing to
do. Am I getting this right?
> +static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
> + struct cgroup *old_cgrp, struct task_struct *task)
> +{
> + /*
> + * cgroup_exit() is called in the copy_process() failure path.
> + * Ignore this case since the task hasn't ran yet, this avoids
> + * trying to poke a half freed task state from generic code.
> + */
> + if (!(task->flags & PF_EXITING))
> + return;
> +
> + perf_cgroup_move(task);
> +}
> +
Those callbacks looks good to me. They certainly alleviate the need for the
hack in cgorup_exit().
Thanks for fixing this.
> +struct cgroup_subsys perf_subsys = {
> + .name = "perf_event",
> + .subsys_id = perf_subsys_id,
> + .create = perf_cgroup_create,
> + .destroy = perf_cgroup_destroy,
> + .exit = perf_cgroup_exit,
> + .attach = perf_cgroup_attach,
> +};
> +#endif /* CONFIG_CGROUP_PERF */
>
>
>
On Tue, 2011-02-08 at 23:31 +0100, Stephane Eranian wrote:
> Peter,
>
> See comments below.
>
>
> On Mon, Feb 7, 2011 at 5:10 PM, Peter Zijlstra <[email protected]> wrote:
> > Compile tested only, depends on the cgroup::exit patch
> >
> > --- linux-2.6.orig/include/linux/perf_event.h
> > +++ linux-2.6/include/linux/perf_event.h
> > @@ -905,6 +929,9 @@ struct perf_cpu_context {
> > struct list_head rotation_list;
> > int jiffies_interval;
> > struct pmu *active_pmu;
> > +#ifdef CONFIG_CGROUP_PERF
> > + struct perf_cgroup *cgrp;
> > +#endif
> > };
> >
> I don't quite understand the motivation for adding cgrp to cpuctx.
>
> > --- linux-2.6.orig/kernel/perf_event.c
> > +++ linux-2.6/kernel/perf_event.c
> > +static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
> > +{
> > + struct perf_cgroup *cgrp_out = cpuctx->cgrp;
> > + if (cgrp_out)
> > + __update_cgrp_time(cgrp_out);
> > +}
> > +
> What's the benefit of this form compared to the original from_task() version?
Answering both questions, I did this so we could still do the
sched_out() while the task has already been flipped to a new cgroup.
Note that both attach and the new exit cgroup_subsys methods are called
after they update the task's cgroup. While they do provide the old
cgroup as an argument, making use of that requires passing that along
which would have been a much more invasive change.
> > + if (mode & PERF_CGROUP_SWOUT)
> > + cpu_ctx_sched_out(cpuctx, EVENT_ALL);
> > +
> > + if (mode & PERF_CGROUP_SWIN) {
> > + cpu_ctx_sched_in(cpuctx, EVENT_ALL, task, 1);
> > + cpuctx->cgrp = perf_cgroup_from_task(task);
> > + }
> > + }
> I think there is a risk on cpuctx->cgrp pointing to stale cgrp information.
> Shouldn't we also set cpuctx->cgrp = NULL on SWOUT?
Yeah, we probably should.
> > +static int __perf_cgroup_move(void *info)
> > +{
> > + struct task_struct *task = info;
> > + perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
> > + return 0;
> > +}
> > +
> > +static void perf_cgroup_move(struct task_struct *task)
> > +{
> > + task_function_call(task, __perf_cgroup_move, task);
> > +}
> > +
> > +static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
> > + struct cgroup *old_cgrp, struct task_struct *task,
> > + bool threadgroup)
> > +{
> > + perf_cgroup_move(task);
> > + if (threadgroup) {
> > + struct task_struct *c;
> > + rcu_read_lock();
> > + list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
> > + perf_cgroup_move(c);
> > + }
> > + rcu_read_unlock();
> > + }
> > +}
> > +
> I suspect my original patch was not necessarily handling the attach completely
> when you move an existing task into a cgroup which was already monitored.
> I think you may have had to wait until a ctxsw. Looks like this callback handles
> this better.
Right, this deals with moving a task into a cgroup that isn't currently
being monitored and its converse, moving it out of a cgroup that is
being monitored.
> Let me make sure I understand the threadgroup iteration, though. I suspect
> this handles the situation where a multi-threaded app is moved into a cgroup
Indeed.
> while there is already cgroup monitoring active. In that case and if we do not
> want to wait until there is at least one ctxsw on all CPUs, then we have to
> check if the other threads are not already running on the other CPUs.If so,
> we need to do a cgroup switch on those CPUs. Otherwise, we have nothing to
> do. Am I getting this right?
Right, so if any of those tasks is currently running, that cpu will be
monitoring their old cgroup, hence we send an IPI to flip cgroups.
> > +static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
> > + struct cgroup *old_cgrp, struct task_struct *task)
> > +{
> > + /*
> > + * cgroup_exit() is called in the copy_process() failure path.
> > + * Ignore this case since the task hasn't ran yet, this avoids
> > + * trying to poke a half freed task state from generic code.
> > + */
> > + if (!(task->flags & PF_EXITING))
> > + return;
> > +
> > + perf_cgroup_move(task);
> > +}
> > +
> Those callbacks looks good to me. They certainly alleviate the need for the
> hack in cgorup_exit().
>
> Thanks for fixing this.
n/p, now all we need to do is get this cgroup_subsys::exit stuff
sorted ;-)
On Wed, Feb 9, 2011 at 10:47 AM, Peter Zijlstra <[email protected]> wrote:
> On Tue, 2011-02-08 at 23:31 +0100, Stephane Eranian wrote:
>> Peter,
>>
>> See comments below.
>>
>>
>> On Mon, Feb 7, 2011 at 5:10 PM, Peter Zijlstra <[email protected]> wrote:
>> > Compile tested only, depends on the cgroup::exit patch
>> >
>> > --- linux-2.6.orig/include/linux/perf_event.h
>> > +++ linux-2.6/include/linux/perf_event.h
>> > @@ -905,6 +929,9 @@ struct perf_cpu_context {
>> > struct list_head rotation_list;
>> > int jiffies_interval;
>> > struct pmu *active_pmu;
>> > +#ifdef CONFIG_CGROUP_PERF
>> > + struct perf_cgroup *cgrp;
>> > +#endif
>> > };
>> >
>> I don't quite understand the motivation for adding cgrp to cpuctx.
>>
>> > --- linux-2.6.orig/kernel/perf_event.c
>> > +++ linux-2.6/kernel/perf_event.c
>> > +static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
>> > +{
>> > + struct perf_cgroup *cgrp_out = cpuctx->cgrp;
>> > + if (cgrp_out)
>> > + __update_cgrp_time(cgrp_out);
>> > +}
>> > +
>> What's the benefit of this form compared to the original from_task() version?
>
> Answering both questions, I did this so we could still do the
> sched_out() while the task has already been flipped to a new cgroup.
> Note that both attach and the new exit cgroup_subsys methods are called
> after they update the task's cgroup. While they do provide the old
> cgroup as an argument, making use of that requires passing that along
> which would have been a much more invasive change.
>
>
>> > + if (mode & PERF_CGROUP_SWOUT)
>> > + cpu_ctx_sched_out(cpuctx, EVENT_ALL);
>> > +
>> > + if (mode & PERF_CGROUP_SWIN) {
>> > + cpu_ctx_sched_in(cpuctx, EVENT_ALL, task, 1);
>> > + cpuctx->cgrp = perf_cgroup_from_task(task);
>> > + }
>> > + }
>> I think there is a risk on cpuctx->cgrp pointing to stale cgrp information.
>> Shouldn't we also set cpuctx->cgrp = NULL on SWOUT?
>
> Yeah, we probably should.
>
>> > +static int __perf_cgroup_move(void *info)
>> > +{
>> > + struct task_struct *task = info;
>> > + perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
>> > + return 0;
>> > +}
>> > +
>> > +static void perf_cgroup_move(struct task_struct *task)
>> > +{
>> > + task_function_call(task, __perf_cgroup_move, task);
>> > +}
>> > +
>> > +static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
>> > + struct cgroup *old_cgrp, struct task_struct *task,
>> > + bool threadgroup)
>> > +{
>> > + perf_cgroup_move(task);
>> > + if (threadgroup) {
>> > + struct task_struct *c;
>> > + rcu_read_lock();
>> > + list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
>> > + perf_cgroup_move(c);
>> > + }
>> > + rcu_read_unlock();
>> > + }
>> > +}
>> > +
>> I suspect my original patch was not necessarily handling the attach completely
>> when you move an existing task into a cgroup which was already monitored.
>> I think you may have had to wait until a ctxsw. Looks like this callback handles
>> this better.
>
> Right, this deals with moving a task into a cgroup that isn't currently
> being monitored and its converse, moving it out of a cgroup that is
> being monitored.
>
>> Let me make sure I understand the threadgroup iteration, though. I suspect
>> this handles the situation where a multi-threaded app is moved into a cgroup
>
> Indeed.
>
>> while there is already cgroup monitoring active. In that case and if we do not
>> want to wait until there is at least one ctxsw on all CPUs, then we have to
>> check if the other threads are not already running on the other CPUs.If so,
>> we need to do a cgroup switch on those CPUs. Otherwise, we have nothing to
>> do. Am I getting this right?
>
> Right, so if any of those tasks is currently running, that cpu will be
> monitoring their old cgroup, hence we send an IPI to flip cgroups.
>
I have built a test case where this would trigger. I launched a multi-threaded
app, and then I move the pid into a cgroup via: echo PID >/cgroup/tests/tasks.
I don't see any perf_cgroup move beyond the PID passed.
I looked at kernel/cgroup.c and I could not find a invocation of
ss->attach() that
would pass threadgroup = true. So I am confused here.
I wonder how the cgroupfs 'echo PID >tasks' interface would make the distinction
between PID and TID. It seems possible to move one thread of a multi-threaded
process into a cgroup but not the others.
>>> while there is already cgroup monitoring active. In that case and if we do not
>>> want to wait until there is at least one ctxsw on all CPUs, then we have to
>>> check if the other threads are not already running on the other CPUs.If so,
>>> we need to do a cgroup switch on those CPUs. Otherwise, we have nothing to
>>> do. Am I getting this right?
>>
>> Right, so if any of those tasks is currently running, that cpu will be
>> monitoring their old cgroup, hence we send an IPI to flip cgroups.
>>
> I have built a test case where this would trigger. I launched a multi-threaded
> app, and then I move the pid into a cgroup via: echo PID >/cgroup/tests/tasks.
> I don't see any perf_cgroup move beyond the PID passed.
>
> I looked at kernel/cgroup.c and I could not find a invocation of
> ss->attach() that
> would pass threadgroup = true. So I am confused here.
>
> I wonder how the cgroupfs 'echo PID >tasks' interface would make the distinction
> between PID and TID. It seems possible to move one thread of a multi-threaded
> process into a cgroup but not the others.
>
You can do this:
# echo PID > cgroup.procs
When the patchset that implements the above feature is accepted. See:
https://lkml.org/lkml/2011/2/7/418
The below commit that confused you is actually a part of the above patchset,
but it sneaked into the kernel accidentally:
commit be367d09927023d081f9199665c8500f69f14d22
Author: Ben Blum <[email protected]>
Date: Wed Sep 23 15:56:31 2009 -0700
cgroups: let ss->can_attach and ss->attach do whole threadgroups at a time
Li,
On Fri, Feb 11, 2011 at 1:55 AM, Li Zefan <[email protected]> wrote:
>>>> while there is already cgroup monitoring active. In that case and if we do not
>>>> want to wait until there is at least one ctxsw on all CPUs, then we have to
>>>> check if the other threads are not already running on the other CPUs.If so,
>>>> we need to do a cgroup switch on those CPUs. Otherwise, we have nothing to
>>>> do. Am I getting this right?
>>>
>>> Right, so if any of those tasks is currently running, that cpu will be
>>> monitoring their old cgroup, hence we send an IPI to flip cgroups.
>>>
>> I have built a test case where this would trigger. I launched a multi-threaded
>> app, and then I move the pid into a cgroup via: echo PID >/cgroup/tests/tasks.
>> I don't see any perf_cgroup move beyond the PID passed.
>>
>> I looked at kernel/cgroup.c and I could not find a invocation of
>> ss->attach() that
>> would pass threadgroup = true. So I am confused here.
>>
>> I wonder how the cgroupfs 'echo PID >tasks' interface would make the distinction
>> between PID and TID. It seems possible to move one thread of a multi-threaded
>> process into a cgroup but not the others.
>>
>
> You can do this:
>
> # echo PID > cgroup.procs
>
> When the patchset that implements the above feature is accepted. See:
>
> https://lkml.org/lkml/2011/2/7/418
>
> The below commit that confused you is actually a part of the above patchset,
> but it sneaked into the kernel accidentally:
>
> commit be367d09927023d081f9199665c8500f69f14d22
> Author: Ben Blum <[email protected]>
> Date: Wed Sep 23 15:56:31 2009 -0700
>
> cgroups: let ss->can_attach and ss->attach do whole threadgroups at a time
>
Ok, that makes more sense now. I wil try with the above patchset applied to
verify this work as expected.
On Fri, Feb 11, 2011 at 10:56 AM, Stephane Eranian <[email protected]> wrote:
> Li,
>
> On Fri, Feb 11, 2011 at 1:55 AM, Li Zefan <[email protected]> wrote:
>>>>> while there is already cgroup monitoring active. In that case and if we do not
>>>>> want to wait until there is at least one ctxsw on all CPUs, then we have to
>>>>> check if the other threads are not already running on the other CPUs.If so,
>>>>> we need to do a cgroup switch on those CPUs. Otherwise, we have nothing to
>>>>> do. Am I getting this right?
>>>>
>>>> Right, so if any of those tasks is currently running, that cpu will be
>>>> monitoring their old cgroup, hence we send an IPI to flip cgroups.
>>>>
>>> I have built a test case where this would trigger. I launched a multi-threaded
>>> app, and then I move the pid into a cgroup via: echo PID >/cgroup/tests/tasks.
>>> I don't see any perf_cgroup move beyond the PID passed.
>>>
>>> I looked at kernel/cgroup.c and I could not find a invocation of
>>> ss->attach() that
>>> would pass threadgroup = true. So I am confused here.
>>>
>>> I wonder how the cgroupfs 'echo PID >tasks' interface would make the distinction
>>> between PID and TID. It seems possible to move one thread of a multi-threaded
>>> process into a cgroup but not the others.
>>>
>>
>> You can do this:
>>
>> # echo PID > cgroup.procs
>>
>> When the patchset that implements the above feature is accepted. See:
>>
>> https://lkml.org/lkml/2011/2/7/418
>>
>> The below commit that confused you is actually a part of the above patchset,
>> but it sneaked into the kernel accidentally:
>>
>> commit be367d09927023d081f9199665c8500f69f14d22
>> Author: Ben Blum <[email protected]>
>> Date: Wed Sep 23 15:56:31 2009 -0700
>>
>> cgroups: let ss->can_attach and ss->attach do whole threadgroups at a time
>>
> Ok, that makes more sense now. I wil try with the above patchset applied to
> verify this work as expected.
>
The above patchset seems to work fine. I add to switch perf_event from
ss->attach() to ss->attach_task() to make it work with multi-threaded apps.
The good thing is that it simplifies the code, because all we have to do now
is simply call perf_cgroup_move(). The caller does the thread iteration.
During this exercise, I found a minor issue in event_sched_out() with the
current perf cgroup patch. I will resubmit the patch incl. Peter's changes
+ some more cleanups and the fix + the update perf cgroup patch.
Thanks.