Re: [PATCH 1/5] kernel-shark: Add new dataloading method to be used by the NumPu interface

["Yordan Karadzhov (VMware)" <y.karadz@gmail.com>]

On 29.05.19 г. 20:36 ч., Steven Rostedt wrote:
> On Thu, 23 May 2019 18:18:08 +0300
> Yordan Karadzhov <ykaradzhov@vmware.com> wrote:
> 
>> The new function loads the content of the trace data file into a
>> table / matrix, made of columns / arrays of data having various integer
>> types. Later those arrays will be wrapped as NumPy arrays.
>>
>> Signed-off-by: Yordan Karadzhov <ykaradzhov@vmware.com>
>> ---
>>   kernel-shark/src/libkshark.c | 155 +++++++++++++++++++++++++++++++++++
>>   kernel-shark/src/libkshark.h |   7 ++
>>   2 files changed, 162 insertions(+)
>>
>> diff --git a/kernel-shark/src/libkshark.c b/kernel-shark/src/libkshark.c
>> index 175279c..ac634fd 100644
>> --- a/kernel-shark/src/libkshark.c
>> +++ b/kernel-shark/src/libkshark.c
>> @@ -957,6 +957,161 @@ ssize_t kshark_load_data_records(struct kshark_context *kshark_ctx,
>>   	return -ENOMEM;
>>   }
>>   
>> +static bool data_matrix_alloc(size_t n_rows, uint64_t **offset_array,
>> +					     uint8_t **cpu_array,
>> +					     uint64_t **ts_array,
>> +					     uint16_t **pid_array,
>> +					     int **event_array)
>> +{
>> +	if (offset_array)
>> +		*offset_array = NULL;
>> +
>> +	if (cpu_array)
>> +		*cpu_array = NULL;
>> +
>> +	if (ts_array)
>> +		*ts_array = NULL;
>> +
>> +	if (pid_array)
>> +		*pid_array = NULL;
>> +
>> +	if (event_array)
>> +		*event_array = NULL;
>> +
>> +	if (offset_array) {
> 
> The way you can do this, is remove the above and have:
> 
>> +		*offset_array = calloc(n_rows, sizeof(**offset_array));
>> +		if (!*offset_array)
> 
> 			return false;
> 
>> +			goto free_all;
>> +	}
>> +
>> +	if (cpu_array) {
>> +		*cpu_array = calloc(n_rows, sizeof(**cpu_array));
>> +		if (!*cpu_array)
> 
> 			goto free_offset;
> 
>> +			goto free_all;
>> +	}
>> +
>> +	if (ts_array) {
>> +		*ts_array = calloc(n_rows, sizeof(**ts_array));
>> +		if (!*ts_array)
> 
> 			goto free_cpu;
> 
>> +			goto free_all;
>> +	}
>> +
>> +	if (pid_array) {
>> +		*pid_array = calloc(n_rows, sizeof(**pid_array));
>> +		if (!*pid_array)
> 
> 			goto free_ts;
> 
>> +			goto free_all;
>> +	}
>> +
>> +	if (event_array) {
>> +		*event_array = calloc(n_rows, sizeof(**event_array));
>> +		if (!*event_array)
> 
> 			goto free_pid;
> 
>> +			goto free_all;
>> +	}
>> +
>> +	return true;
>> +
> 
> We can have a helper function:
> 
> static inline void free_ptr(void *ptr)
> {
> 	if (ptr)
> 		free(*(void **)ptr);
> }
> 
> free_pid:
> 	free_ptr(pid_array);
> free_ts:
> 	free_ptr(ts_array);
> free_cpu:
> 	free_ptr(cpu_array);
> free_offset:
> 	free_ptr(offset_array);
> 
> Then have the print here.
> 
>> + free_all:
>> +	fprintf(stderr, "Failed to allocate memory during data loading.\n");
> 
> 	return false;
> 
> This is the way it's usually done in the Linux kernel.
> 

Great! Really elegant solution.


>> +
>> +	if (offset_array)
>> +		free(*offset_array);
>> +
>> +	if (cpu_array)
>> +		free(*cpu_array);
>> +
>> +	if (ts_array)
>> +		free(*ts_array);
>> +
>> +	if (pid_array)
>> +		free(*pid_array);
>> +
>> +	if (event_array)
>> +		free(*event_array);
>> +
>> +	return false;
>> +}
>> +
>> +/**
>> + * @brief Load the content of the trace data file into a table / matrix made
>> + *	  of columns / arrays of data. The user is responsible for freeing the
>> + *	  elements of the outputted array
>> + *
>> + * @param kshark_ctx: Input location for the session context pointer.
>> + * @param offset_array: Output location for the array of record offsets.
>> + * @param cpu_array: Output location for the array of CPU Ids.
>> + * @param ts_array: Output location for the array of timestamps.
>> + * @param pid_array: Output location for the array of Process Ids.
>> + * @param event_array: Output location for the array of Event Ids.
>> + *
>> + * @returns The size of the outputted arrays in the case of success, or a
>> + *	    negative error code on failure.
>> + */
>> +size_t kshark_load_data_matrix(struct kshark_context *kshark_ctx,
>> +			       uint64_t **offset_array,
>> +			       uint8_t **cpu_array,
>> +			       uint64_t **ts_array,
>> +			       uint16_t **pid_array,
>> +			       int **event_array)
>> +{
>> +	enum rec_type type = REC_ENTRY;
>> +	struct rec_list **rec_list;
>> +	size_t count, total = 0;
>> +	bool status;
>> +	int n_cpus;
>> +
>> +	total = get_records(kshark_ctx, &rec_list, type);
>> +	if (total < 0)
>> +		goto fail;
>> +
>> +	n_cpus = tracecmd_cpus(kshark_ctx->handle);
>> +
>> +	status = data_matrix_alloc(total, offset_array,
>> +					  cpu_array,
>> +					  ts_array,
>> +					  pid_array,
>> +					  event_array);
> 
> BTW, have you looked into how much memory this takes up in a large
> trace?

One record takes 24 bytes. So it will allocate total*24 bytes (or less 
if some of the input arguments is NULL).


> 
>> +	if (!status)
>> +		goto fail_free;
>> +
>> +	for (count = 0; count < total; count++) {
>> +		int next_cpu;
>> +
>> +		next_cpu = pick_next_cpu(rec_list, n_cpus, type);
>> +		if (next_cpu >= 0) {
>> +			struct kshark_entry *e = &rec_list[next_cpu]->entry;
> 
> Hmm, this looks like we are taking an address of a field and then
> freeing it down below. Looking at the definition of rec_list, this is
> currently OK. But this coding style is not robust because of the tight
> dependency to how rec_list is defined. If that ever changes it will be
> hard to find code like this to update it.
> 
> A more robust way to do this is:
> 
> 			struct rec_list *rec = rec_list[next_cpu];
> 			struct kshark_entry *e = &rec->entry;
> 
>> +
>> +			if (offset_array)
>> +				(*offset_array)[count] = e->offset;
>> +
>> +			if (cpu_array)
>> +				(*cpu_array)[count] = e->cpu;
>> +
>> +			if (ts_array)
>> +				(*ts_array)[count] = e->ts;
>> +
>> +			if (pid_array)
>> +				(*pid_array)[count] = e->pid;
>> +
>> +			if (event_array)
>> +				(*event_array)[count] = e->event_id;
>> +
>> +			rec_list[next_cpu] = rec_list[next_cpu]->next;
> 
> Then here:
> 
> 			free(rec);
> 
> That way there's not a dependency here with the data structure layout
> of rec_list and the freeing of the entry.
> 
> -- Steve
> 

Thanks a lot! Sending updated version.
Y.

> 
>> +			free(e);
>> +		}
>> +	}
>> +
>> +	/* There should be no entries left in rec_list. */
>> +	free_rec_list(rec_list, n_cpus, type);
>> +	return total;
>> +
>> + fail_free:
>> +	free_rec_list(rec_list, n_cpus, type);
>> +
>> + fail:
>> +	fprintf(stderr, "Failed to allocate memory during data
>> loading.\n");
>> +	return -ENOMEM;
>> +}
>> +
>>   static const char *kshark_get_latency(struct tep_handle *pe,
>>   				      struct tep_record *record)
>>   {
>> diff --git a/kernel-shark/src/libkshark.h
>> b/kernel-shark/src/libkshark.h index c218b61..92ade41 100644
>> --- a/kernel-shark/src/libkshark.h
>> +++ b/kernel-shark/src/libkshark.h
>> @@ -149,6 +149,13 @@ ssize_t kshark_load_data_entries(struct
>> kshark_context *kshark_ctx, ssize_t kshark_load_data_records(struct
>> kshark_context *kshark_ctx, struct tep_record ***data_rows);
>>   
>> +size_t kshark_load_data_matrix(struct kshark_context *kshark_ctx,
>> +			       uint64_t **offset_array,
>> +			       uint8_t **cpu_array,
>> +			       uint64_t **ts_array,
>> +			       uint16_t **pid_array,
>> +			       int **event_array);
>> +
>>   ssize_t kshark_get_task_pids(struct kshark_context *kshark_ctx, int
>> **pids);
>>   void kshark_close(struct kshark_context *kshark_ctx);
>