perf_events: add cgroup support (v8)

kernel/cgroup.c +7 -0

@@ -4259,8 +4261,20 @@ void cgroup_exit(struct task_struct *tsk, int run_callbacks)
 
        /* Reassign the task to the init_css_set. */
        task_lock(tsk);
+       /*
+        * we mask interrupts to prevent:
+        * - timer tick to cause event rotation which
+        *   could schedule back in cgroup events after
+        *   they were switched out by perf_cgroup_sched_out()
+        *
+        * - preemption which could schedule back in cgroup events

Quoted 11 lines · Stephane Eranian

kernel/cgroup.c +7 -0

@@ -4259,8 +4261,20 @@ void cgroup_exit(struct task_struct *tsk, int run_callbacks)
 
         /* Reassign the task to the init_css_set. */
         task_lock(tsk);
+       /*
+        * we mask interrupts to prevent:
+        * - timer tick to cause event rotation which
+        *   could schedule back in cgroup events after
+        *   they were switched out by perf_cgroup_sched_out()
+        *
+        * - preemption which could schedule back in cgroup events

kernel/cgroup.c void cgroup_exit(struct task_struct *tsk, int run_callbacks) · line 4280

Quoted 10 lines

kernel/cgroup.c +5 -0

+        */
+       local_irq_save(flags);
+       perf_cgroup_sched_out(tsk);
         cg = tsk->cgroups;
         tsk->cgroups = &init_css_set;
+       perf_cgroup_sched_in(tsk);
+       local_irq_restore(flags);
         task_unlock(tsk);
         if (cg)
                 put_css_set_taskexit(cg);

So you too need a callback on cgroup change there.. Li, Paul, any chance we can fix this cgroup_subsys::exit callback? The scheduler code needs to do funny thing because its in the wrong place as well.

Quoted 17 lines · Stephane Eranian

kernel/cgroup.c +11 -0

@@ -4259,8 +4261,20 @@ void cgroup_exit(struct task_struct *tsk, int run_callbacks)
 
        /* Reassign the task to the init_css_set. */
        task_lock(tsk);
+       /*
+        * we mask interrupts to prevent:
+        * - timer tick to cause event rotation which
+        *   could schedule back in cgroup events after
+        *   they were switched out by perf_cgroup_sched_out()
+        *
+        * - preemption which could schedule back in cgroup events
+        */
+       local_irq_save(flags);
+       perf_cgroup_sched_out(tsk);
        cg = tsk->cgroups;
        tsk->cgroups = &init_css_set;
+       perf_cgroup_sched_in(tsk);

kernel/cgroup.c void cgroup_exit(struct task_struct *tsk, int run_callbacks) · line 4280

Quoted 4 lines

kernel/cgroup.c +1 -0

+       local_irq_restore(flags);
        task_unlock(tsk);
        if (cg)
                put_css_set_taskexit(cg); 

So you too need a callback on cgroup change there.. Li, Paul, any chance we can fix this cgroup_subsys::exit callback? The scheduler code needs to do funny thing because its in the wrong place as well.

Quoted 18 lines · Peter Zijlstra

Quoted 17 lines · Stephane Eranian

kernel/cgroup.c +11 -0

@@ -4259,8 +4261,20 @@ void cgroup_exit(struct task_struct *tsk, int run_callbacks)
 
        /* Reassign the task to the init_css_set. */
        task_lock(tsk);
+       /*
+        * we mask interrupts to prevent:
+        * - timer tick to cause event rotation which
+        *   could schedule back in cgroup events after
+        *   they were switched out by perf_cgroup_sched_out()
+        *
+        * - preemption which could schedule back in cgroup events
+        */
+       local_irq_save(flags);
+       perf_cgroup_sched_out(tsk);
        cg = tsk->cgroups;
        tsk->cgroups = &init_css_set;
+       perf_cgroup_sched_in(tsk);

Quoted 8 lines

kernel/cgroup.c void cgroup_exit(struct task_struct *tsk, int run_callbacks) · line 4280

Quoted 4 lines

kernel/cgroup.c +1 -0

+       local_irq_restore(flags);
        task_unlock(tsk);
        if (cg)
                put_css_set_taskexit(cg); 

So you too need a callback on cgroup change there.. Li, Paul, any chance we can fix this cgroup_subsys::exit callback? The scheduler code needs to do funny thing because its in the wrong place as well.

cgroup guys? Shall I just fix this exit thing since the only user seems to be the scheduler and now perf for both of which its unfortunate at best?a

Balbir, memcontrol.c uses pre_destroy(), I pose that using this method is broken per definition since it makes the cgroup empty notification void.

Quoted 28 lines · Peter Zijlstra

Quoted 26 lines · Peter Zijlstra

Quoted 11 lines · Stephane Eranian

kernel/cgroup.c +7 -0

@@ -4259,8 +4261,20 @@ void cgroup_exit(struct task_struct *tsk, int run_callbacks)
 
        /* Reassign the task to the init_css_set. */
        task_lock(tsk);
+       /*
+        * we mask interrupts to prevent:
+        * - timer tick to cause event rotation which
+        *   could schedule back in cgroup events after
+        *   they were switched out by perf_cgroup_sched_out()
+        *
+        * - preemption which could schedule back in cgroup events

kernel/cgroup.c void cgroup_exit(struct task_struct *tsk, int run_callbacks) · line 4280

Quoted 10 lines

kernel/cgroup.c +5 -0

+        */
+       local_irq_save(flags);
+       perf_cgroup_sched_out(tsk);
        cg = tsk->cgroups;
        tsk->cgroups = &init_css_set;
+       perf_cgroup_sched_in(tsk);
+       local_irq_restore(flags);
        task_unlock(tsk);
        if (cg)
                put_css_set_taskexit(cg); 

So you too need a callback on cgroup change there.. Li, Paul, any chance we can fix this cgroup_subsys::exit callback? The scheduler code needs to do funny thing because its in the wrong place as well.

Quoted 3 lines

cgroup guys? Shall I just fix this exit thing since the only user seems to be the scheduler and now perf for both of which its unfortunate at best?

Are you suggesting that the cgroup_exit on task_exit notification should be pulled out?

Quoted 4 lines

Balbir, memcontrol.c uses pre_destroy(), I pose that using this method is broken per definition since it makes the cgroup empty notification void.

We use pre_destroy() to reclaim, so that delete/rmdir() will be able to clean up the node/group. I am not sure what you mean by it makes the empty notification void and why pre_destroy() is broken?

Quoted 34 lines · Balbir Singh

Quoted 31 lines · Peter Zijlstra

Quoted 18 lines · Peter Zijlstra

Quoted 17 lines · Stephane Eranian

kernel/cgroup.c +11 -0

@@ -4259,8 +4261,20 @@ void cgroup_exit(struct task_struct *tsk, int run_callbacks)
 
        /* Reassign the task to the init_css_set. */
        task_lock(tsk);
+       /*
+        * we mask interrupts to prevent:
+        * - timer tick to cause event rotation which
+        *   could schedule back in cgroup events after
+        *   they were switched out by perf_cgroup_sched_out()
+        *
+        * - preemption which could schedule back in cgroup events
+        */
+       local_irq_save(flags);
+       perf_cgroup_sched_out(tsk);
        cg = tsk->cgroups;
        tsk->cgroups = &init_css_set;
+       perf_cgroup_sched_in(tsk);

Quoted 8 lines

kernel/cgroup.c void cgroup_exit(struct task_struct *tsk, int run_callbacks) · line 4280

Quoted 4 lines

kernel/cgroup.c +1 -0

+       local_irq_restore(flags);
        task_unlock(tsk);
        if (cg)
                put_css_set_taskexit(cg); 

So you too need a callback on cgroup change there.. Li, Paul, any chance we can fix this cgroup_subsys::exit callback? The scheduler code needs to do funny thing because its in the wrong place as well.

cgroup guys? Shall I just fix this exit thing since the only user seems to be the scheduler and now perf for both of which its unfortunate at best?

Quoted 2 lines

Are you suggesting that the cgroup_exit on task_exit notification should be pulled out?

No, just fixed. The callback as it exists isn't useful and leads to hacks like the above.

Quoted 9 lines

Quoted 5 lines

Balbir, memcontrol.c uses pre_destroy(), I pose that using this method is broken per definition since it makes the cgroup empty notification void.

We use pre_destroy() to reclaim, so that delete/rmdir() will be able to clean up the node/group. I am not sure what you mean by it makes the empty notification void and why pre_destroy() is broken?

A quick look at the code looked like it could return -EBUSY (and other errors), in that case the rmdir of the empty cgroup will fail.

Therefore it can happen that after the last task is removed, and we get the notification that the cgroup is empty, and we attempt the rmdir we will fail.

This again means that all such notification handlers must poll state, which is ridiculous.

Quoted 2 lines

No, just fixed. The callback as it exists isn't useful and leads to hacks like the above.

Make the ::exit method act like ::attach, it is after all very nearly the same thing.

 void cgroup_exit(struct task_struct *tsk, int run_callbacks) {
-       int i;
       struct css_set *cg;
+       int i;

-       if (run_callbacks && need_forkexit_callback) {
-               /*
-                * modular subsystems can't use callbacks, so no need to lock
-                * the subsys array
-                */
-               for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
-                       struct cgroup_subsys *ss = subsys[i];
-                       if (ss->exit)
-                               ss->exit(ss, tsk);
-               }
-       }
+       mutex_lock(&cgroup_mutex);

Quoted 3 lines · Peter Zijlstra

Make the ::exit method act like ::attach, it is after all very nearly the same thing.

The major difference between attach and exit is that the former is only triggered in response to user cgroup-management action, whereas the latter is triggered whenever a task exits, even if cgroups aren't set up.

Quoted 8 lines

void cgroup_exit(struct task_struct *tsk, int run_callbacks) {

kernel/cgroup.c +1 -3

-       int i;
       struct css_set *cg;
+       int i;
 
-       if (run_callbacks && need_forkexit_callback) {
-               /*

kernel/cgroup.c void cgroup_post_fork(struct task_struct · line 4236

Quoted 10 lines

kernel/cgroup.c +1 -9

-                * modular subsystems can't use callbacks, so no need to lock
-                * the subsys array
-                */
-               for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
-                       struct cgroup_subsys *ss = subsys[i];
-                       if (ss->exit)
-                               ss->exit(ss, tsk);
-               }
-       }
+       mutex_lock(&cgroup_mutex);

NACK - cgroup_mutex is way too heavy to take in the task exit path. We'll need to find some other way to fix this if it's really needed. task->alloc_lock is also normally a valid thing to synchronize against cgroup moves, but I'd have to look at the exit path to see if it's still valid there.

Quoted 5 lines · Paul Menage

Quoted 3 lines · Peter Zijlstra

Make the ::exit method act like ::attach, it is after all very nearly the same thing.

Quoted 4 lines

The major difference between attach and exit is that the former is only triggered in response to user cgroup-management action, whereas the latter is triggered whenever a task exits, even if cgroups aren't set up.

And the major likeness is that they both migrate a task from one cgroup to another. You cannot simply ignore that.

If maybe you're like respond more often than about once every two months I might actually care what you think.

Quoted 11 lines · Peter Zijlstra

Quoted 9 lines · Paul Menage

Quoted 3 lines · Peter Zijlstra

Make the ::exit method act like ::attach, it is after all very nearly the same thing.

The major difference between attach and exit is that the former is only triggered in response to user cgroup-management action, whereas the latter is triggered whenever a task exits, even if cgroups aren't set up.

Quoted 2 lines

And the major likeness is that they both migrate a task from one cgroup to another. You cannot simply ignore that.

True, and the exit path for cgroups has always been a bit fuzzy - that was kind of inherited from cpusets, where this worked out OK, but the CPU subsystem has more interesting requirements.

An important semantic difference between attach and exit is that in the exit case the destination is always the root, and the task in question is going to be exiting before doing anything else interesting. So it should be possible to optimize that path a lot compared to the regular attach (many things related to resource usage can be ignored, since the task won't have time to actually use any non-trivial amount of resources).

Quoted 3 lines

If maybe you're like respond more often than about once every two months I might actually care what you think.

Yes, sadly since switching groups at Google, cgroups has become pretty much just a spare-time activity for me

- but that in itself doesn't
automatically invalidate my opinion when I do have time to respond. It's still the case that cgroup_mutex is an incredibly heavyweight mutex that has no business being in the task exit path. Or do you just believe that ad hominem is a valid style of argument?

How about if the exit callback was moved before the preceding task_unlock()? Since I think the scheduler is still the only user of the exit callback, redefining the locking semantics should be fine.

Quoted 3 lines

How about if the exit callback was moved before the preceding task_unlock()? Since I think the scheduler is still the only user of the exit callback, redefining the locking semantics should be fine.

Like the below? Both the perf and sched exit callback are fine with being called under task_lock afaict, but I haven't actually ran with lockdep enabled to see if I missed something.

kernel/cgroup.c +4 -0

+			}
+		}
+	}
 	task_unlock(tsk);
+
 	if (cg)
 		put_css_set_taskexit(cg);
 }

Since both sched and perf won't use the 2 args @cgrp and @old_cgrp, don't bother to change the ->exit interface?

And here I thought Google was starting to understand what community participation meant.. is there anybody you know who can play a more active role in the whole cgroup thing?

Like the below? Both the perf and sched exit callback are fine with being called under task_lock afaict, but I haven't actually ran with

lockdep enabled to see if I missed something.

I also pondered doing the cgroup exit from __put_task_struct() but that had another problem iirc.

---
Index: linux-2.6/include/linux/cgroup.h
===================================================================

include/linux/cgroup.h +2 -1

@@ -474,7 +474,8 @@ struct cgroup_subsys {
                        struct cgroup *old_cgrp, struct task_struct *tsk,
                        bool threadgroup);
        void (*fork)(struct cgroup_subsys *ss, struct task_struct *task);
-       void (*exit)(struct cgroup_subsys *ss, struct task_struct *task);
+       void (*exit)(struct cgroup_subsys *ss, struct cgroup *cgrp,
+                       struct cgroup *old_cgrp, struct task_struct *task);
        int (*populate)(struct cgroup_subsys *ss,
                        struct cgroup *cgrp);
        void (*post_clone)(struct cgroup_subsys *ss, struct cgroup *cgrp);

Index: linux-2.6/kernel/cgroup.c
===================================================================

kernel/cgroup.c +18 -13

@@ -4230,20 +4230,8 @@ void cgroup_post_fork(struct task_struct
  */
  void cgroup_exit(struct task_struct *tsk, int run_callbacks)
  {
-       int i;
        struct css_set *cg;
-
-       if (run_callbacks && need_forkexit_callback) {
-               /*
-                * modular subsystems can't use callbacks, so no need to lock
-                * the subsys array
-                */
-               for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
-                       struct cgroup_subsys *ss = subsys[i];
-                       if (ss->exit)
-                               ss->exit(ss, tsk);
-               }
-       }
+       int i;
 
        /*
         * Unlink from the css_set task list if necessary.

@@ -4261,7 +4249,24 @@ void cgroup_exit(struct task_struct *tsk
        task_lock(tsk);
        cg = tsk->cgroups;
        tsk->cgroups = &init_css_set;
+
+       if (run_callbacks && need_forkexit_callback) {
+               /*
+                * modular subsystems can't use callbacks, so no need to lock
+                * the subsys array
+                */
+               for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
+                       struct cgroup_subsys *ss = subsys[i];
+                       if (ss->exit) {
+                               struct cgroup *old_cgrp =
+                                       rcu_dereference_raw(cg->subsys[i])->cgroup;
+                               struct cgroup *cgrp = task_cgroup(tsk, i);
+                               ss->exit(ss, cgrp, old_cgrp, tsk);
+                       }
+               }
+       }
        task_unlock(tsk);
+
        if (cg)
                put_css_set_taskexit(cg);
  }

Index: linux-2.6/kernel/sched.c
===================================================================

kernel/sched.c +0 -3

@@ -606,9 +606,6 @@ static inline struct task_group *task_gr
        struct task_group *tg;
        struct cgroup_subsys_state *css;
 
-       if (p->flags & PF_EXITING)
-               return &root_task_group;
-
        css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
                        lockdep_is_held(&task_rq(p)->lock));
        tg = container_of(css, struct task_group, css);

@@ -9081,7 +9078,8 @@ cpu_cgroup_attach(struct cgroup_subsys *
  }
 

 static void

kernel/sched.c +2 -1

-cpu_cgroup_exit(struct cgroup_subsys *ss, struct task_struct *task)
+cpu_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.

Quoted 2 lines · Balbir Singh

Quoted 2 lines

No, just fixed. The callback as it exists isn't useful and leads to hacks like the above.

---
Subject: cgroup: Fix cgroup_subsys::exit callback
From: Peter Zijlstra <[email protected]>
Date: Mon Feb 07 17:02:20 CET 2011

Make the ::exit method act like ::attach, it is after all very nearly the same thing.

Signed-off-by Peter Zijlstra <[email protected]>

LKML-Reference: <new-submission>
---
Index: linux-2.6/include/linux/cgroup.h
===================================================================

Quoted 30 lines · Balbir Singh

Quoted 28 lines · Peter Zijlstra

Quoted 26 lines · Peter Zijlstra

Quoted 11 lines · Stephane Eranian

kernel/cgroup.c +7 -0

@@ -4259,8 +4261,20 @@ void cgroup_exit(struct task_struct *tsk, int run_callbacks)
 
         /* Reassign the task to the init_css_set. */
         task_lock(tsk);
+       /*
+        * we mask interrupts to prevent:
+        * - timer tick to cause event rotation which
+        *   could schedule back in cgroup events after
+        *   they were switched out by perf_cgroup_sched_out()
+        *
+        * - preemption which could schedule back in cgroup events

kernel/cgroup.c void cgroup_exit(struct task_struct *tsk, int run_callbacks) · line 4280

Quoted 10 lines

kernel/cgroup.c +5 -0

+        */
+       local_irq_save(flags);
+       perf_cgroup_sched_out(tsk);
         cg = tsk->cgroups;
         tsk->cgroups = &init_css_set;
+       perf_cgroup_sched_in(tsk);
+       local_irq_restore(flags);
         task_unlock(tsk);
         if (cg)
                 put_css_set_taskexit(cg);

So you too need a callback on cgroup change there.. Li, Paul, any chance we can fix this cgroup_subsys::exit callback? The scheduler code needs to do funny thing because its in the wrong place as well.

Quoted 6 lines

Quoted 3 lines

cgroup guys? Shall I just fix this exit thing since the only user seems to be the scheduler and now perf for both of which its unfortunate at best?

Are you suggesting that the cgroup_exit on task_exit notification should be pulled out?

No, just fixed. The callback as it exists isn't useful and leads to hacks like the above.

Quoted 8 lines

Quoted 4 lines

Balbir, memcontrol.c uses pre_destroy(), I pose that using this method is broken per definition since it makes the cgroup empty notification void.

We use pre_destroy() to reclaim, so that delete/rmdir() will be able to clean up the node/group. I am not sure what you mean by it makes the empty notification void and why pre_destroy() is broken?

A quick look at the code looked like it could return -EBUSY (and other errors), in that case the rmdir of the empty cgroup will fail.

Therefore it can happen that after the last task is removed, and we get the notification that the cgroup is empty, and we attempt the rmdir we will fail.

This again means that all such notification handlers must poll state, which is ridiculous.

Quoted 3 lines

This again means that all such notification handlers must poll state, which is ridiculous.

Not necessarily - we could make it that a failed rmdir() sets a bit that causes a notification again once the final refcount is dropped again on the cgroup.

Also, the core cgroup code should deal with the cgroup refcount, not individual controllers.

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
===================================================================

include/linux/cgroup.h +1 -0

@@ -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
===================================================================

include/linux/cgroup_subsys.h +4 -0

@@ -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
===================================================================

include/linux/perf_event.h +29 -2

@@ -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

include/linux/perf_event.h +1 -1

@@ -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
===================================================================

init/Kconfig +0 -0

@@ -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).

init/Kconfig +10 -0

+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
===================================================================

kernel/cgroup.c +23 -0

@@ -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
===================================================================

kernel/perf_event.c +329 -2

@@ -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;

kernel/perf_event.c +18 -0

@@ -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

kernel/perf_event.c +1 -1

-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

kernel/perf_event.c +0 -0

@@ -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;
       }

kernel/perf_event.c +36 -0

@@ -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,

kernel/perf_event.c +31 -6

@@ -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);

kernel/perf_event.c +0 -0

@@ -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;

kernel/perf_event.c +5 -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,

kernel/perf_event.c +20 -11

@@ -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);
 }

kernel/perf_event.c +9 -3

@@ -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 ||

kernel/perf_event.c +2 -2

@@ -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) ||

kernel/perf_event.c +108 -3

@@ -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 */

kernel/perf_event.c +32 -0

+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.
+                */

kernel/perf_event.c +17 -0

+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);

kernel/perf_event.c +4 -0

+static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
+               struct cgroup *old_cgrp, struct task_struct *task)
+{
+       /*

kernel/perf_event.c +9 -0

+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.

include/linux/perf_event.h struct perf_cpu_context { · line 936

Quoted 13 lines · Peter Zijlstra

Compile tested only, depends on the cgroup::exit patch

include/linux/perf_event.h +3 -0

@@ -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.

kernel/perf_event.c static inline u64 perf_clock(void) · line 245

Quoted 9 lines

kernel/perf_event.c +7 -0

+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?

kernel/perf_event.c static inline u64 perf_clock(void) · line 322

Quoted 8 lines

kernel/perf_event.c +8 -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?

Quoted 23 lines

kernel/perf_event.c +23 -0

+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);
+               }

kernel/perf_event.c static inline u64 perf_clock(void) · line 336

Quoted 4 lines

kernel/perf_event.c +4 -0

+               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?

Quoted 4 lines

kernel/perf_event.c +4 -0

+static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
+               struct cgroup *old_cgrp, struct task_struct *task)
+{
+       /*

kernel/perf_event.c static int __init perf_event_sysfs_init( · line 7343

Quoted 10 lines

kernel/perf_event.c +10 -0

+        * 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.

Quoted 21 lines · Stephane Eranian

Peter,

See comments below.

include/linux/perf_event.h struct perf_cpu_context { · line 936

Quoted 13 lines · Peter Zijlstra

Compile tested only, depends on the cgroup::exit patch

include/linux/perf_event.h +3 -0

@@ -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.

Quoted 10 lines

kernel/perf_event.c static inline u64 perf_clock(void) · line 245

Quoted 9 lines

kernel/perf_event.c +7 -0

+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.

Quoted 10 lines

Quoted 8 lines

+                       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.

Quoted 23 lines

Quoted 23 lines

kernel/perf_event.c +23 -0

+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);
+               }

Quoted 8 lines

kernel/perf_event.c

Quoted 4 lines

+               rcu_read_unlock();
+       }

kernel/perf_event.c +2 -0

+}
+

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.

Quoted 2 lines

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.

Quoted 5 lines

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.

Quoted 9 lines

Quoted 5 lines

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.

Quoted 13 lines

Quoted 9 lines

Quoted 5 lines

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.

Quoted 8 lines

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,

Quoted 33 lines · Li Zefan

Quoted 13 lines

Quoted 9 lines

Quoted 5 lines

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.

Quoted 8 lines

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:

Quoted 6 lines

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.