Re: [PATCH] oprofile, perf: Use per-cpu framework

[William Cohen <wcohen@redhat.com>]

On 06/22/2012 10:42 AM, Robert Richter wrote:
> Will,
> 
> On 15.06.12 17:19:54, Will Deacon wrote:
>> On Fri, Jun 15, 2012 at 05:16:35PM +0100, Robert Richter wrote:
>>> On 15.06.12 16:43:59, Will Deacon wrote:
>>>> Given both the lack of feedback and the trivial nature of this patch would
>>>> you be able to pick this up please? I'd rather not push it via the ARM tree
>>>> as the change is under drivers/.
>>>
>>> Will, I will apply it. Sorry for not responding earlier.
> 
> I just realized I wrote a patch some time ago that introduces per_cpu
> variables, see below. Please give it a try to, I will then queue it
> for v3.6.
> 
> Your patch should go upstream anyway to make it into stable trees.
> 
> Thanks,
> 
> -Robert

Hi Robert,

I was able to build a linux-3.5.0-rc5+ kernel with this patch and the following troublesome config entry enabled:

CONFIG_DEBUG_PER_CPU_MAPS=y

The resulting kernel ran and oprofile functioned fine on a compulab trimslice machine.

-Will

> 
> 
> From f8bbfd7d28303967ca4e8597de9bdc9bf8b197e7 Mon Sep 17 00:00:00 2001
> From: Robert Richter <robert.richter@amd.com>
> Date: Thu, 23 Feb 2012 17:07:06 +0100
> Subject: [PATCH] oprofile, perf: Use per-cpu framework
> 
> This changes oprofile_perf.c to use the per-cpu framework.
> 
> Using the per-cpu framework should avoid error like the following:
> 
>  arch/arm/oprofile/../../../drivers/oprofile/oprofile_perf.c:28:28: error: variably modified 'perf_events' at file scope
> 
> Reported-by: William Cohen <wcohen@redhat.com>
> Cc: Will Deacon <will.deacon@arm.com>
> Signed-off-by: Robert Richter <robert.richter@amd.com>
> ---
>  drivers/oprofile/oprofile_perf.c |   23 +++++++++++------------
>  1 files changed, 11 insertions(+), 12 deletions(-)
> 
> diff --git a/drivers/oprofile/oprofile_perf.c b/drivers/oprofile/oprofile_perf.c
> index efc4b7f..f3cfa0b 100644
> --- a/drivers/oprofile/oprofile_perf.c
> +++ b/drivers/oprofile/oprofile_perf.c
> @@ -1,5 +1,6 @@
>  /*
>   * Copyright 2010 ARM Ltd.
> + * Copyright 2012 Advanced Micro Devices, Inc., Robert Richter
>   *
>   * Perf-events backend for OProfile.
>   */
> @@ -25,7 +26,7 @@ static int oprofile_perf_enabled;
>  static DEFINE_MUTEX(oprofile_perf_mutex);
>  
>  static struct op_counter_config *counter_config;
> -static struct perf_event **perf_events[NR_CPUS];
> +static DEFINE_PER_CPU(struct perf_event **, perf_events);
>  static int num_counters;
>  
>  /*
> @@ -38,7 +39,7 @@ static void op_overflow_handler(struct perf_event *event,
>  	u32 cpu = smp_processor_id();
>  
>  	for (id = 0; id < num_counters; ++id)
> -		if (perf_events[cpu][id] == event)
> +		if (per_cpu(perf_events, cpu)[id] == event)
>  			break;
>  
>  	if (id != num_counters)
> @@ -74,7 +75,7 @@ static int op_create_counter(int cpu, int event)
>  {
>  	struct perf_event *pevent;
>  
> -	if (!counter_config[event].enabled || perf_events[cpu][event])
> +	if (!counter_config[event].enabled || per_cpu(perf_events, cpu)[event])
>  		return 0;
>  
>  	pevent = perf_event_create_kernel_counter(&counter_config[event].attr,
> @@ -91,18 +92,18 @@ static int op_create_counter(int cpu, int event)
>  		return -EBUSY;
>  	}
>  
> -	perf_events[cpu][event] = pevent;
> +	per_cpu(perf_events, cpu)[event] = pevent;
>  
>  	return 0;
>  }
>  
>  static void op_destroy_counter(int cpu, int event)
>  {
> -	struct perf_event *pevent = perf_events[cpu][event];
> +	struct perf_event *pevent = per_cpu(perf_events, cpu)[event];
>  
>  	if (pevent) {
>  		perf_event_release_kernel(pevent);
> -		perf_events[cpu][event] = NULL;
> +		per_cpu(perf_events, cpu)[event] = NULL;
>  	}
>  }
>  
> @@ -257,12 +258,12 @@ void oprofile_perf_exit(void)
>  
>  	for_each_possible_cpu(cpu) {
>  		for (id = 0; id < num_counters; ++id) {
> -			event = perf_events[cpu][id];
> +			event = per_cpu(perf_events, cpu)[id];
>  			if (event)
>  				perf_event_release_kernel(event);
>  		}
>  
> -		kfree(perf_events[cpu]);
> +		kfree(per_cpu(perf_events, cpu));
>  	}
>  
>  	kfree(counter_config);
> @@ -277,8 +278,6 @@ int __init oprofile_perf_init(struct oprofile_operations *ops)
>  	if (ret)
>  		return ret;
>  
> -	memset(&perf_events, 0, sizeof(perf_events));
> -
>  	num_counters = perf_num_counters();
>  	if (num_counters <= 0) {
>  		pr_info("oprofile: no performance counters\n");
> @@ -298,9 +297,9 @@ int __init oprofile_perf_init(struct oprofile_operations *ops)
>  	}
>  
>  	for_each_possible_cpu(cpu) {
> -		perf_events[cpu] = kcalloc(num_counters,
> +		per_cpu(perf_events, cpu) = kcalloc(num_counters,
>  				sizeof(struct perf_event *), GFP_KERNEL);
> -		if (!perf_events[cpu]) {
> +		if (!per_cpu(perf_events, cpu)) {
>  			pr_info("oprofile: failed to allocate %d perf events "
>  					"for cpu %d\n", num_counters, cpu);
>  			ret = -ENOMEM;