Re: [PATCH] perfcounters: Support for ftrace event records sampling

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Fri, 2009-08-07 at 01:25 +0200, Frederic Weisbecker wrote:
> This patch brings the kernel side support for ftrace event record
> sampling.
> 
> A new counter attribute is added: PERF_SAMPLE_TP_RECORD which requests
> ftrace events record sampling.
> 

> +	PERF_SAMPLE_TP_RECORD			= 1U << 10,

I'd really want this thing called PERF_SAMPLE_RAW

> -	PERF_SAMPLE_MAX = 1U << 10,		/* non-ABI */
> +	PERF_SAMPLE_MAX = 1U << 11,		/* non-ABI */
>  };
>  
>  /*
> @@ -413,6 +414,11 @@ struct perf_callchain_entry {
>  	__u64				ip[PERF_MAX_STACK_DEPTH];
>  };
>  
> +struct perf_tracepoint_record {
> +	int				size;
> +	char				*record;
> +};

Which would make this:

struct perf_raw_record {
	u32	size;
	void 	*data;
};

>  struct task_struct;
>  
>  /**
> @@ -681,6 +687,7 @@ struct perf_sample_data {
>  	struct pt_regs			*regs;
>  	u64				addr;
>  	u64				period;
> +	void				*private;
>  };

might as well make that struct perf_raw_record *raw;

> @@ -649,5 +617,99 @@ __attribute__((section("_ftrace_events"))) event_##call = {		\
>  
>  #include TRACE_INCLUDE(TRACE_INCLUDE_FILE)
>  
> +/*
> + * Define the insertion callback to profile events
> + *
> + * The job is very similar to ftrace_raw_event_<call> except that we don't
> + * insert in the ring buffer but in a perf counter.
> + *
> + * static void ftrace_profile_<call>(proto)
> + * {
> + *	struct ftrace_data_offsets_<call> __maybe_unused __data_offsets;
> + *	struct ftrace_event_call *event_call = &event_<call>;
> + *	extern void perf_tpcounter_event(int, u64, u64, void *, int);
> + *	struct ftrace_raw_##call *entry;
> + *	u64 __addr = 0, __count = 1;
> + *	unsigned long irq_flags;
> + *	int __entry_size;
> + *	int __data_size;
> + *	int pc;
> + *
> + *	local_save_flags(irq_flags);
> + *	pc = preempt_count();
> + *
> + *	__data_size = ftrace_get_offsets_<call>(&__data_offsets, args);
> + *	__entry_size = __data_size + sizeof(*entry);
> + *
> + *	do {
> + *		char raw_data[__entry_size]; <- allocate our sample in the stack
> + *		struct trace_entry *ent;
> + *
> + *		entry = (struct ftrace_raw_<call> *)raw_data;
> + *		ent = &entry->ent;
> + *		tracing_generic_entry_update(ent, irq_flags, pc);
> + *		ent->type = event_call->id;
> + *
> + *		<tstruct> <- do some jobs with dynamic arrays
> + *
> + *		<assign>  <- affect our values
> + *
> + *		perf_tpcounter_event(event_call->id, __addr, __count, entry,
> + *			     __entry_size);  <- submit them to perf counter
> + *	} while (0);
> + *
> + * }
> + */
> +
> +#ifdef CONFIG_EVENT_PROFILE
> +
> +#undef __perf_addr
> +#define __perf_addr(a) __addr = (a)
> +
> +#undef __perf_count
> +#define __perf_count(c) __count = (c)
> +
> +#undef TRACE_EVENT
> +#define TRACE_EVENT(call, proto, args, tstruct, assign, print)		\
> +static void ftrace_profile_##call(proto)				\
> +{									\
> +	struct ftrace_data_offsets_##call __maybe_unused __data_offsets;\
> +	struct ftrace_event_call *event_call = &event_##call;		\
> +	extern void perf_tpcounter_event(int, u64, u64, void *, int);	\
> +	struct ftrace_raw_##call *entry;				\
> +	u64 __addr = 0, __count = 1;					\
> +	unsigned long irq_flags;					\
> +	int __entry_size;						\
> +	int __data_size;						\
> +	int pc;								\
> +									\
> +	local_save_flags(irq_flags);					\
> +	pc = preempt_count();						\
> +									\
> +	__data_size = ftrace_get_offsets_##call(&__data_offsets, args); \
> +	__entry_size = __data_size + sizeof(*entry);			\
> +									\
> +	do {								\
> +		char raw_data[__entry_size];				\
> +		struct trace_entry *ent;				\
> +									\
> +		entry = (struct ftrace_raw_##call *)raw_data;		\
> +		ent = &entry->ent;					\
> +		tracing_generic_entry_update(ent, irq_flags, pc);	\
> +		ent->type = event_call->id;				\
> +									\
> +		tstruct							\
> +									\
> +		{ assign; }						\
> +									\
> +		perf_tpcounter_event(event_call->id, __addr, __count, entry,\
> +			     __entry_size);				\
> +	} while (0);							\
> +									\
> +}

ok, so the one concern I have here is that the data needs to fit on the
stack. What if someone puts a large string in the data?

> +
> +#include TRACE_INCLUDE(TRACE_INCLUDE_FILE)
> +#endif /* CONFIG_EVENT_PROFILE */
> +
>  #undef _TRACE_PROFILE_INIT
>  
> diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
> index 06d210c..93f4312 100644
> --- a/kernel/perf_counter.c
> +++ b/kernel/perf_counter.c
> @@ -2646,6 +2646,7 @@ static void perf_counter_output(struct perf_counter *counter, int nmi,
>  		u64 counter;
>  	} group_entry;
>  	struct perf_callchain_entry *callchain = NULL;
> +	struct perf_tracepoint_record *tp;
>  	int callchain_size = 0;
>  	u64 time;
>  	struct {
> @@ -2714,6 +2715,11 @@ static void perf_counter_output(struct perf_counter *counter, int nmi,
>  			header.size += sizeof(u64);
>  	}
>  
> +	if (sample_type & PERF_SAMPLE_TP_RECORD) {
> +		tp = data->private;
> +		header.size += tp->size;
> +	}
> +
>  	ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
>  	if (ret)
>  		return;
> @@ -2777,6 +2783,9 @@ static void perf_counter_output(struct perf_counter *counter, int nmi,
>  		}
>  	}
>  
> +	if (sample_type & PERF_SAMPLE_TP_RECORD)
> +		perf_output_copy(&handle, tp->record, tp->size);
> +
>  	perf_output_end(&handle);
>  }

You seem to fail to round up to a multiple of u64 somewhere along the
line, that'll mess things up as events are supposed to be u64 aligned.


> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index 6da0992..90c9808 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -412,6 +412,7 @@ static void create_counter(int counter, int cpu, pid_t pid)
>  	if (call_graph)
>  		attr->sample_type	|= PERF_SAMPLE_CALLCHAIN;
>  
> +
>  	attr->mmap		= track;
>  	attr->comm		= track;
>  	attr->inherit		= (cpu < 0) && inherit;

Do we really need that extra whitespace?