Re: [Cbe-oss-dev] [PATCH] Cell OProfile: SPU mutex lock fix

[Carl Love]

On Fri, 2008-04-04 at 08:38 +0200, Arnd Bergmann wrote: 
 On Wednesday 02 April 2008, Carl Love wrote:
  On Wed, 2008-04-02 at 07:21 +0200, Arnd Bergmann wrote:
   On Tuesday 25 March 2008, Carl Love wrote:
This patch fixes a bug in the code that records the SPU data and
context switches.  The buffer_mutex lock must be held when the
kernel is adding data to the buffer between the kernel and the
OProfile daemon.  The lock is not being held in the current code
base.  This patch fixes the bug using work queues.  The data to 
be passed to the daemon is caputured by the interrupt handler.  
The workqueue function is invoked to grab the buffer_mutex lock
and add the data to the buffer.  
   
   So what was the exact bug you're fixing with this? There was no
   buffer_mutex before, so why do you need it now? Can't this be a
   spinlock so you can get it from interrupt context instead of
   using a workqueue?
  
  The generic OProfile code defines a mutex lock, called buffer_mutex, to
  protect the kernel/daemon data buffer from being writen by the kernal
  and simultaneously read by the Daemon.  When adding a PPU sample the
  oprofile routine  oprofile_add_ext_sample(pc, regs, i, is_kernel) is
  called from the interrupt context to request the sample be stored.  The
  generic oprofile code takes care of passing the data to a non interrupt
  context where the mutex lock is held and the necessary sequence of data
  is written into the kernel/daemon data buffer.  However, OProfile does
  not have any built in functions for handling the SPU.  Hence, we have to
  implement the code to capture the data in the interrupt context, pass it
  to a non interrupt context and put it into the buffer.  This was not
  done correctly in the original implementation.  Specifically, the mutex
  lock was not being held.  
 
 Ok, I see.
 
 However, I'm pretty sure that the switch notification does not get
 called from an atomic context, so you don't need a workqueue for
 bringing that into a process context. Doing the context switch
 notification directly from the scheduler sounds much better regarding
 the impact on the measurement.

Our first thought to fix the bug was to just grab the mutex lock when
adding the switch notification data to the queue.  The kernel gives us
an oops message saying something along the line of could not call mutex
lock in interrupt context.  Hence we had to go to work queues so we
could access the lock outside of the SPU switch notification context.

Secondly, it is my understanding that if the schedule_work() call tries
to queue the same work function multiple times the subsequent requests
are dropped.  Thus we were not able to pass the context switch data as
part of the schedule work requests.  This forced us to have an array to
store the data for each SPU.   
 
   Never put extern statements in the implementation, they describe the
   interface between two parts of the code and should be inside of a
   common header.
   
   Why do you want to have your own workqueue instead of using the
   global one?
  
  It is important that the data get context switch data get recorded as
  quickly as possible to avoid dropping data unnecessarily.  The PC
  counter data for each SPU is ignored until the context switch record is
  put into the kernel/daemon buffer.  The API documentation says that
  using a private workqueue has better performance then using the global
  workqueue.  There is a comment in the code about this, perhaps it is not
  clear enough.
 
 This sounds like an unrelated bug in the implementation. The PC
 data should *not* be ignored in any case. As long as the records
 get stored in the right order, everything should be fine here.

Until the OProfile sees the context switch record, it does not know what
to do with the PC samples and just drops them.  The thought was using a
private work queue might help get the context switch records processed a
little earlier.  It probably doesn't make that much difference.  I can
just use the generic work queue.  
 
 
   This looks like you want to use a delayed_work rather than building your
   own out of hrtimer and work. Is there any point why you want to use
   an hrtimer?
  
  The current implementation uses the hrtimer to schedule when to read the
  trace buffer the next time.  This patch does not change how the
  scheduling of the buffer reads is done.  Yes, you could change the
  implementation to use workqueues instead.  If you feel that it is better
  to use the workqueue then we could make that change.  Not sure that
  making that change in this bug fix patch is appropriate.  I would need
  to create a second patch for that change.
 
 I would guess that the change from hrtimer to delayed_workqueue is
 smaller than the current patch changing from hrtimer to hrtimer plus
 workqueue, so I would prefer to have only one changeset.
 
 Since the timer only causes statistical data collection anyway, delaying
 it a bit should 

Re: [Cbe-oss-dev] [PATCH] Cell OProfile: SPU mutex lock fix

[Arnd Bergmann]

On Tuesday 08 April 2008, Carl Love wrote:
 On Fri, 2008-04-04 at 08:38 +0200, Arnd Bergmann wrote: 
 Our first thought to fix the bug was to just grab the mutex lock when
 adding the switch notification data to the queue.  The kernel gives us
 an oops message saying something along the line of could not call mutex
 lock in interrupt context.  Hence we had to go to work queues so we
 could access the lock outside of the SPU switch notification context.

By the time the notifier is called, the kernel is certainly not
in an atomic context. Maybe you were nesting the mutex lock inside
of your own spinlock?

 Secondly, it is my understanding that if the schedule_work() call tries
 to queue the same work function multiple times the subsequent requests
 are dropped.  Thus we were not able to pass the context switch data as
 part of the schedule work requests.  This forced us to have an array to
 store the data for each SPU.   

The way that work queues are designed, you embed the struct work in the
data you pass down, so that you are guaranteed to execute the work struct
exactly once per data element and you don't need any global pointers.

  I would guess that the change from hrtimer to delayed_workqueue is
  smaller than the current patch changing from hrtimer to hrtimer plus
  workqueue, so I would prefer to have only one changeset.
  
  Since the timer only causes statistical data collection anyway, delaying
  it a bit should not have any negative effect on the accuracy of the
  measurement, unlike delaying the context switch notification.
 
 The goal is to be able to support very high sampling rates (small
 periods).  The schedule_delayed_work() is based on jiffies which I
 believe is 1/250 for this machine.  This only gives millisecond
 resolution.  The goal is for the users to be able to specify a time
 period of 60,000 cycles or less then 20 micro second sampling periods
 when the real high resolution timers are available.  We can't achieve
 the desired sampling rates with the schedule_dealyed_work() function.

You actually can't get anywhere close to that resolution if you do your
data collection in a work queue, because under high load (which is what
the only time when measuring really gets interesting) the work queue
is likely to be delayed by a few jiffies!

If you rely on high resolution timer precision, you need to look
at the performance counters from inside the timer function, and deal
with the problem of the work queue not being called for a number of
timer events.

 
 Oprofile provides nice clean interfaces for recording kernel/user
 switches and CPU data recording.  This is all that was needed by any
 architecture until CELL came along. With CELL, we now have need to add
 processor data plus SPU data to the queue.  The buffer_mutex variable
 and the add_event_entry() were not visible outside of the OProfile
 driver code.  The original SPU support added add_event_entry() to the
 include/linux/oprofile.h file.  We can add the buffer_mutex as well
 since there is now a need to access both of these.  
 
 I have been looking to see how I could create a generic oprofile routine
 which could take the data.  The routine would still have to work from an
 interrupt context, so it will need to store the data and call a work
 queue function.  The function would need to know how much data will be
 needed, thus you would probably need to statically allocate data or use
 a list and malloc the data as needed.  I don't really want to have to
 malloc data from an interrupt context.  List management adds additional
 overhead.  It would be possible to have an init function that you could
 call at startup time telling it how much memory you need, in this case
 we could allocate a buffer the size of spu_info (defined below) at
 startup time.  The call could pass an array to the OProfile routine that
 would put the data into the buffer and call the work function.  We still
 have to allocate the storage, it does clean up the arch specific code.
 Not sure if this really buys us much.  There is more copying of data
 i.e. more overhead.  Not convinced the OProfile maintainers would accept
 anything I have thought of so far.
 
 Any suggestions?

My first attempt to do this would be to add this to the oprofile/cpu_buffer.c
infrastructure. Basically extend the add_sample() function to have
helpers you can call from the spu code to put entries into the per-cpu
buffer of the CPU that happens to execute the code at the time.

add_sample() can already be called from an atomic context since it uses
its own buffer, and it uses a clever ring buffer to get around the
need for locking against the event_buffer functions.
Only event_buffer needs the mutex, so it's best to leave that out
of the architecture code running at interrupt time altogether.

  An ideal driver should not have *any* global variables at all, but store
  all data in the (reference counted) objects it is dealing with, or
  just on the stack while it's processing 

Re: [Cbe-oss-dev] [PATCH] Cell OProfile: SPU mutex lock fix

[Arnd Bergmann]

On Wednesday 02 April 2008, Carl Love wrote:
 On Wed, 2008-04-02 at 07:21 +0200, Arnd Bergmann wrote:
  On Tuesday 25 March 2008, Carl Love wrote:
   This patch fixes a bug in the code that records the SPU data and
   context switches.  The buffer_mutex lock must be held when the
   kernel is adding data to the buffer between the kernel and the
   OProfile daemon.  The lock is not being held in the current code
   base.  This patch fixes the bug using work queues.  The data to 
   be passed to the daemon is caputured by the interrupt handler.  
   The workqueue function is invoked to grab the buffer_mutex lock
   and add the data to the buffer.  
  
  So what was the exact bug you're fixing with this? There was no
  buffer_mutex before, so why do you need it now? Can't this be a
  spinlock so you can get it from interrupt context instead of
  using a workqueue?
 
 The generic OProfile code defines a mutex lock, called buffer_mutex, to
 protect the kernel/daemon data buffer from being writen by the kernal
 and simultaneously read by the Daemon.  When adding a PPU sample the
 oprofile routine  oprofile_add_ext_sample(pc, regs, i, is_kernel) is
 called from the interrupt context to request the sample be stored.  The
 generic oprofile code takes care of passing the data to a non interrupt
 context where the mutex lock is held and the necessary sequence of data
 is written into the kernel/daemon data buffer.  However, OProfile does
 not have any built in functions for handling the SPU.  Hence, we have to
 implement the code to capture the data in the interrupt context, pass it
 to a non interrupt context and put it into the buffer.  This was not
 done correctly in the original implementation.  Specifically, the mutex
 lock was not being held.  

Ok, I see.

However, I'm pretty sure that the switch notification does not get
called from an atomic context, so you don't need a workqueue for
bringing that into a process context. Doing the context switch
notification directly from the scheduler sounds much better regarding
the impact on the measurement.

  Never put extern statements in the implementation, they describe the
  interface between two parts of the code and should be inside of a
  common header.
  
  Why do you want to have your own workqueue instead of using the
  global one?
 
 It is important that the data get context switch data get recorded as
 quickly as possible to avoid dropping data unnecessarily.  The PC
 counter data for each SPU is ignored until the context switch record is
 put into the kernel/daemon buffer.  The API documentation says that
 using a private workqueue has better performance then using the global
 workqueue.  There is a comment in the code about this, perhaps it is not
 clear enough.

This sounds like an unrelated bug in the implementation. The PC
data should *not* be ignored in any case. As long as the records
get stored in the right order, everything should be fine here.


  This looks like you want to use a delayed_work rather than building your
  own out of hrtimer and work. Is there any point why you want to use
  an hrtimer?
 
 The current implementation uses the hrtimer to schedule when to read the
 trace buffer the next time.  This patch does not change how the
 scheduling of the buffer reads is done.  Yes, you could change the
 implementation to use workqueues instead.  If you feel that it is better
 to use the workqueue then we could make that change.  Not sure that
 making that change in this bug fix patch is appropriate.  I would need
 to create a second patch for that change.

I would guess that the change from hrtimer to delayed_workqueue is
smaller than the current patch changing from hrtimer to hrtimer plus
workqueue, so I would prefer to have only one changeset.

Since the timer only causes statistical data collection anyway, delaying
it a bit should not have any negative effect on the accuracy of the
measurement, unlike delaying the context switch notification.

   -static DEFINE_SPINLOCK(buffer_lock);
   +extern struct mutex buffer_mutex;
   +extern struct workqueue_struct *oprofile_spu_wq;
   +extern int calls_to_record_switch;
   +
  
  Again, public interfaces need to go to a header file, and should
  have a name that identifies the interface. buffer_mutex is
  certainly not a suitable name for a kernel-wide global variable!
 
 As stated earlier, the generic OProfile code defines the variable
 buffer_mutex.  Changing the name in the generic OProfile code is
 beyond the scope of this patch.

Ok, didn't see that the name was already part of the main oprofile
driver. However, this makes it even worse: you are accessing data
structures that are clearly not meant to be shared with architecture
code. The fact that it was not declared in a global header file should
have told you that.

I think you should instead add a function to drivers/oprofile/buffer_sync.c
that takes care of moving the data to the common buffer under the right
mutex_lock.

  
static 

Re: [Cbe-oss-dev] [PATCH] Cell OProfile: SPU mutex lock fix

[Carl Love]

On Wed, 2008-04-02 at 07:21 +0200, Arnd Bergmann wrote:
 On Tuesday 25 March 2008, Carl Love wrote:
  This patch fixes a bug in the code that records the SPU data and
  context switches.  The buffer_mutex lock must be held when the
  kernel is adding data to the buffer between the kernel and the
  OProfile daemon.  The lock is not being held in the current code
  base.  This patch fixes the bug using work queues.  The data to 
  be passed to the daemon is caputured by the interrupt handler.  
  The workqueue function is invoked to grab the buffer_mutex lock
  and add the data to the buffer.  
 
 So what was the exact bug you're fixing with this? There was no
 buffer_mutex before, so why do you need it now? Can't this be a
 spinlock so you can get it from interrupt context instead of
 using a workqueue?

The generic OProfile code defines a mutex lock, called buffer_mutex, to
protect the kernel/daemon data buffer from being writen by the kernal
and simultaneously read by the Daemon.  When adding a PPU sample the
oprofile routine  oprofile_add_ext_sample(pc, regs, i, is_kernel) is
called from the interrupt context to request the sample be stored.  The
generic oprofile code takes care of passing the data to a non interrupt
context where the mutex lock is held and the necessary sequence of data
is written into the kernel/daemon data buffer.  However, OProfile does
not have any built in functions for handling the SPU.  Hence, we have to
implement the code to capture the data in the interrupt context, pass it
to a non interrupt context and put it into the buffer.  This was not
done correctly in the original implementation.  Specifically, the mutex
lock was not being held.  

Writing data to the OProfile buffer consists of a sequence of items.
For example when writing an SPU entry, first comes the escape code so
the daemon knows this is a new entry.  The next item is the SPU context
switch code which says the data which will follow is the information
about a new context.  There is a different code to identify the data as
an address sample.  Finally the data about the SPU context switch is
entered into the buffer.  The issue is the OProfile daemon is read all
of the entire sequence of items then process the data.  Without the
mutex lock, the daemon may read part of the sequence try to process it
before everything is written into the buffer.  When the daemon reads
again, it doesn't see the escape code as the first item and isn't smart
enough to realize it is part of a previous sequence.  The generic
OProfile code defines the mutex lock and calls it buffer_mutex.  The
OProfile kernel/daemon API uses the mutex lock. The mutex lock can only
be held in a non interrupt context.  The current implementation uses a
spin lock to make sure the kernel writes each sequence if items into the
buffer but since the API does not use a spin lock we have no way to
prevent the daemon from reading the buffer until the entire sequence of
items has been written to the buffer.  Hence the need to hold the
buffer_mutex lock which prevents the daemon from accessing the buffer.

 
  Index: linux-2.6.25-rc4/arch/powerpc/oprofile/cell/spu_profiler.c
  ===
  --- linux-2.6.25-rc4.orig/arch/powerpc/oprofile/cell/spu_profiler.c
  +++ linux-2.6.25-rc4/arch/powerpc/oprofile/cell/spu_profiler.c
  @@ -16,6 +16,7 @@
   #include linux/smp.h
   #include linux/slab.h
   #include asm/cell-pmu.h
  +#include linux/workqueue.h
   #include pr_util.h
   
 
 Please keep #include statements in alphabetical order, with all linux/ files
 before the asm/ files.
 
   #define TRACE_ARRAY_SIZE 1024
  @@ -32,9 +33,19 @@ static unsigned int profiling_interval;
   
   #define SPU_PC_MASK 0x
   
  +/* The generic OProfile code uses the buffer_mutex to protect the buffer
  + * between the kernel and the daemon.  The SPU code needs to use the buffer
  + * to ensure that the kernel SPU writes complete as a single block before
  + * being consumed by the daemon.
  + */
  +extern struct mutex buffer_mutex;
  +
   static DEFINE_SPINLOCK(sample_array_lock);
   unsigned long sample_array_lock_flags;
   
  +struct work_struct spu_record_wq;
  +extern struct workqueue_struct *oprofile_spu_wq;
  +
   void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int 
  cycles_reset)
   {
  unsigned long ns_per_cyc;
 
 Never put extern statements in the implementation, they describe the
 interface between two parts of the code and should be inside of a
 common header.
 
 Why do you want to have your own workqueue instead of using the
 global one?
 
  @@ -123,14 +134,14 @@ static int cell_spu_pc_collection(int cp
  return entry;
   }
   
  -
  -static enum hrtimer_restart profile_spus(struct hrtimer *timer)
  -{
  -   ktime_t kt;
  +static void profile_spus_record_samples (struct work_struct *ws) {
  +   /* This routine is called via schedule_work() to record the
  +* spu data.  It must be run 

Re: [Cbe-oss-dev] [PATCH] Cell OProfile: SPU mutex lock fix

[Arnd Bergmann]

On Tuesday 25 March 2008, Carl Love wrote:
 This patch fixes a bug in the code that records the SPU data and
 context switches.  The buffer_mutex lock must be held when the
 kernel is adding data to the buffer between the kernel and the
 OProfile daemon.  The lock is not being held in the current code
 base.  This patch fixes the bug using work queues.  The data to 
 be passed to the daemon is caputured by the interrupt handler.  
 The workqueue function is invoked to grab the buffer_mutex lock
 and add the data to the buffer.  

So what was the exact bug you're fixing with this? There was no
buffer_mutex before, so why do you need it now? Can't this be a
spinlock so you can get it from interrupt context instead of
using a workqueue?

 Index: linux-2.6.25-rc4/arch/powerpc/oprofile/cell/spu_profiler.c
 ===
 --- linux-2.6.25-rc4.orig/arch/powerpc/oprofile/cell/spu_profiler.c
 +++ linux-2.6.25-rc4/arch/powerpc/oprofile/cell/spu_profiler.c
 @@ -16,6 +16,7 @@
  #include linux/smp.h
  #include linux/slab.h
  #include asm/cell-pmu.h
 +#include linux/workqueue.h
  #include pr_util.h
  

Please keep #include statements in alphabetical order, with all linux/ files
before the asm/ files.

  #define TRACE_ARRAY_SIZE 1024
 @@ -32,9 +33,19 @@ static unsigned int profiling_interval;
  
  #define SPU_PC_MASK   0x
  
 +/* The generic OProfile code uses the buffer_mutex to protect the buffer
 + * between the kernel and the daemon.  The SPU code needs to use the buffer
 + * to ensure that the kernel SPU writes complete as a single block before
 + * being consumed by the daemon.
 + */
 +extern struct mutex buffer_mutex;
 +
  static DEFINE_SPINLOCK(sample_array_lock);
  unsigned long sample_array_lock_flags;
  
 +struct work_struct spu_record_wq;
 +extern struct workqueue_struct *oprofile_spu_wq;
 +
  void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int 
 cycles_reset)
  {
   unsigned long ns_per_cyc;

Never put extern statements in the implementation, they describe the
interface between two parts of the code and should be inside of a
common header.

Why do you want to have your own workqueue instead of using the
global one?

 @@ -123,14 +134,14 @@ static int cell_spu_pc_collection(int cp
   return entry;
  }
  
 -
 -static enum hrtimer_restart profile_spus(struct hrtimer *timer)
 -{
 - ktime_t kt;
 +static void profile_spus_record_samples (struct work_struct *ws) {
 + /* This routine is called via schedule_work() to record the
 +  * spu data.  It must be run in a normal kernel mode to
 +  * grab the OProfile mutex lock.
 +  */
   int cpu, node, k, num_samples, spu_num;
  
 - if (!spu_prof_running)
 - goto stop;
 + mutex_lock(buffer_mutex);
  
   for_each_online_cpu(cpu) {
   if (cbe_get_hw_thread_id(cpu))
 @@ -170,6 +181,20 @@ static enum hrtimer_restart profile_spus
   smp_wmb();  /* insure spu event buffer updates are written */
   /* don't want events intermingled... */
  
 + mutex_unlock(buffer_mutex);
 +}
 +
 +static enum hrtimer_restart profile_spus(struct hrtimer *timer)
 +{
 + ktime_t kt;
 +
 +
 + if (!spu_prof_running)
 + goto stop;
 +
 + /* schedule the funtion to record the data */
 + schedule_work(spu_record_wq);
 +
   kt = ktime_set(0, profiling_interval);
   if (!spu_prof_running)
   goto stop;

This looks like you want to use a delayed_work rather than building your
own out of hrtimer and work. Is there any point why you want to use
an hrtimer?

 -static DEFINE_SPINLOCK(buffer_lock);
 +extern struct mutex buffer_mutex;
 +extern struct workqueue_struct *oprofile_spu_wq;
 +extern int calls_to_record_switch;
 +

Again, public interfaces need to go to a header file, and should
have a name that identifies the interface. buffer_mutex is
certainly not a suitable name for a kernel-wide global variable!

  static DEFINE_SPINLOCK(cache_lock);
  static int num_spu_nodes;
 +
  int spu_prof_num_nodes;
  int last_guard_val[MAX_NUMNODES * 8];
 +int cnt_swtch_processed_flag[MAX_NUMNODES * 8];
 +
 +struct spus_profiling_code_data_s {
 + int num_spu_nodes;
 + struct work_struct spu_prof_code_wq;
 +} spus_profiling_code_data;
 +
 +struct spu_context_switch_data_s {
 + struct spu *spu;
 + unsigned long spu_cookie;
 + unsigned long app_dcookie;
 + unsigned int offset;
 + unsigned long objectId;
 + int valid_entry;
 +} spu_context_switch_data;

I don't understand what these variables are really doing, but
having e.g. just one spu_context_switch_data for all the SPUs
doesn't seem to make much sense. What happens when two SPUs do
a context switch at the same time?

 +int calls_to_record_switch = 0;
 +int record_spu_start_flag = 0;
 +
 +struct spus_cntxt_sw_data_s {
 + int num_spu_nodes;
 + struct spu_context_switch_data_s spu_data[MAX_NUMNODES * 8];
 + struct 

[Cbe-oss-dev] [PATCH] Cell OProfile: SPU mutex lock fix

[Carl Love]

This patch fixes a bug in the code that records the SPU data and
context switches.  The buffer_mutex lock must be held when the
kernel is adding data to the buffer between the kernel and the
OProfile daemon.  The lock is not being held in the current code
base.  This patch fixes the bug using work queues.  The data to 
be passed to the daemon is caputured by the interrupt handler.  
The workqueue function is invoked to grab the buffer_mutex lock
and add the data to the buffer.  

Signed-off-by: Carl Love [EMAIL PROTECTED]


Index: linux-2.6.25-rc4/arch/powerpc/oprofile/cell/spu_profiler.c
===
--- linux-2.6.25-rc4.orig/arch/powerpc/oprofile/cell/spu_profiler.c
+++ linux-2.6.25-rc4/arch/powerpc/oprofile/cell/spu_profiler.c
@@ -16,6 +16,7 @@
 #include linux/smp.h
 #include linux/slab.h
 #include asm/cell-pmu.h
+#include linux/workqueue.h
 #include pr_util.h
 
 #define TRACE_ARRAY_SIZE 1024
@@ -32,9 +33,19 @@ static unsigned int profiling_interval;
 
 #define SPU_PC_MASK 0x
 
+/* The generic OProfile code uses the buffer_mutex to protect the buffer
+ * between the kernel and the daemon.  The SPU code needs to use the buffer
+ * to ensure that the kernel SPU writes complete as a single block before
+ * being consumed by the daemon.
+ */
+extern struct mutex buffer_mutex;
+
 static DEFINE_SPINLOCK(sample_array_lock);
 unsigned long sample_array_lock_flags;
 
+struct work_struct spu_record_wq;
+extern struct workqueue_struct *oprofile_spu_wq;
+
 void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int 
cycles_reset)
 {
unsigned long ns_per_cyc;
@@ -123,14 +134,14 @@ static int cell_spu_pc_collection(int cp
return entry;
 }
 
-
-static enum hrtimer_restart profile_spus(struct hrtimer *timer)
-{
-   ktime_t kt;
+static void profile_spus_record_samples (struct work_struct *ws) {
+   /* This routine is called via schedule_work() to record the
+* spu data.  It must be run in a normal kernel mode to
+* grab the OProfile mutex lock.
+*/
int cpu, node, k, num_samples, spu_num;
 
-   if (!spu_prof_running)
-   goto stop;
+   mutex_lock(buffer_mutex);
 
for_each_online_cpu(cpu) {
if (cbe_get_hw_thread_id(cpu))
@@ -170,6 +181,20 @@ static enum hrtimer_restart profile_spus
smp_wmb();  /* insure spu event buffer updates are written */
/* don't want events intermingled... */
 
+   mutex_unlock(buffer_mutex);
+}
+
+static enum hrtimer_restart profile_spus(struct hrtimer *timer)
+{
+   ktime_t kt;
+
+
+   if (!spu_prof_running)
+   goto stop;
+
+   /* schedule the funtion to record the data */
+   schedule_work(spu_record_wq);
+
kt = ktime_set(0, profiling_interval);
if (!spu_prof_running)
goto stop;
@@ -209,6 +234,10 @@ int start_spu_profiling(unsigned int cyc
spu_prof_running = 1;
hrtimer_start(timer, kt, HRTIMER_MODE_REL);
 
+   /* setup the workqueue for recording the SPU data */
+   INIT_WORK(spu_record_wq,
+ profile_spus_record_samples);
+
return 0;
 }
 
Index: linux-2.6.25-rc4/arch/powerpc/oprofile/cell/spu_task_sync.c
===
--- linux-2.6.25-rc4.orig/arch/powerpc/oprofile/cell/spu_task_sync.c
+++ linux-2.6.25-rc4/arch/powerpc/oprofile/cell/spu_task_sync.c
@@ -27,15 +27,44 @@
 #include linux/numa.h
 #include linux/oprofile.h
 #include linux/spinlock.h
+#include linux/workqueue.h
 #include pr_util.h
 
 #define RELEASE_ALL 
 
-static DEFINE_SPINLOCK(buffer_lock);
+extern struct mutex buffer_mutex;
+extern struct workqueue_struct *oprofile_spu_wq;
+extern int calls_to_record_switch;
+
 static DEFINE_SPINLOCK(cache_lock);
 static int num_spu_nodes;
+
 int spu_prof_num_nodes;
 int last_guard_val[MAX_NUMNODES * 8];
+int cnt_swtch_processed_flag[MAX_NUMNODES * 8];
+
+struct spus_profiling_code_data_s {
+   int num_spu_nodes;
+   struct work_struct spu_prof_code_wq;
+} spus_profiling_code_data;
+
+struct spu_context_switch_data_s {
+   struct spu *spu;
+   unsigned long spu_cookie;
+   unsigned long app_dcookie;
+   unsigned int offset;
+   unsigned long objectId;
+   int valid_entry;
+} spu_context_switch_data;
+
+int calls_to_record_switch = 0;
+int record_spu_start_flag = 0;
+
+struct spus_cntxt_sw_data_s {
+   int num_spu_nodes;
+   struct spu_context_switch_data_s spu_data[MAX_NUMNODES * 8];
+   struct work_struct spu_cntxt_work;
+} spus_cntxt_sw_data;
 
 /* Container for caching information about an active SPU task. */
 struct cached_info {
@@ -44,6 +73,8 @@ struct cached_info {
struct kref cache_ref;
 };
 
+struct workqueue_struct *oprofile_spu_wq;
+
 static struct cached_info *spu_info[MAX_NUMNODES * 8];
 
 static void destroy_cached_info(struct kref *kref)
@@ -283,39 +314,90