/* \param xdf pointer of a valid xdf file
*
* Allocate all the buffers and temporary objects needed for the transfer
*/
static int alloc_transfer_objects(struct xdf* xdf)
{
unsigned int samsize;
xdf->sample_size = samsize = compute_sample_size(xdf, 1);
xdf->filerec_size = compute_sample_size(xdf, 0) * xdf->ns_per_rec;
if ( !(xdf->convdata = malloc(xdf->numch*sizeof(*(xdf->convdata))))
|| !(xdf->batch = malloc(xdf->nbatch*sizeof(*(xdf->batch))))
|| !(xdf->buff = malloc(xdf->ns_per_rec * samsize))
|| !(xdf->backbuff = malloc(xdf->ns_per_rec * samsize))
|| !(xdf->tmpbuff[0] = malloc(xdf->ns_per_rec * 8))
|| !(xdf->tmpbuff[1] = malloc(xdf->ns_per_rec * 8)) ) {
return -1;
}
return 0;
}
/*
* Arguments:
* -i input file. Default: perf.data.
*/
int
main(int argc, char **argv)
{
int ret, c;
char *input_fname = "perf.data";
char *output_fname = "perf.data.manicured";
char *optval = NULL;
perf_file_header f_header, f_header_manicured;
while ((c = getopt (argc, argv, "b:e:i:o:s:")) != -1)
{
switch(c)
{
case 'b':
begin_time = parse_timestamp_and_exit_on_error(optarg, "begin");
break;
case 'e':
end_time = parse_timestamp_and_exit_on_error(optarg, "end");
break;
case 'i':
input_fname = optarg;
break;
case 'o':
output_fname = optarg;
break;
case 's':
user_base_time = parse_timestamp_and_exit_on_error(optarg, "start");
break;
case '?':
default:
usage(argv[0]);
exit(-1);
}
}
printf("Begin timestamp: %" PRIu64 " \n"
"End timestamp: %" PRIu64 " \n"
"Start (of program) timestamp: %" PRIu64 " \n",
begin_time, end_time, user_base_time);
if(user_base_time == 0 && (begin_time != 0))
printf("Warning: zero starting timestamp provided. Your begin and end timestamps will not be correctly "
"calibrated to perf timestamps.\n");
if(begin_time < user_base_time || end_time < user_base_time)
{
printf("Your begin or end timestamps are smaller than the starting timestamp. Cannot proceed.\n");
usage(argv[0]);
exit(-1);
}
/* Let's reset begin and end timestamps to be relative to the start-of-program timestamp,
* so we don't have to perform this computation on every event.
*/
begin_time -= user_base_time;
end_time -= user_base_time;
int ifd = open(input_fname, O_RDONLY);
if(ifd == -1)
{
fprintf(stderr, "Could not open %s: %s\n", input_fname, strerror(errno));
usage(argv[0]);
exit(-1);
}
int ofd = open(output_fname, O_CREAT | O_RDWR, S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP);
if(ofd == -1)
{
fprintf(stderr, "Could not open %s: %s\n", output_fname, strerror(errno));
usage(argv[0]);
exit(-1);
}
ret = check_and_copy_header(ifd, ofd, &f_header);
if(ret)
exit(-1);
printf("Successful header check...\n");
/*
* Sanity check that perf.data was written cleanly; data size is
* initialized to 0 and updated only if the on_exit function is run.
* If data size is still 0 then the file contains only partial
* information. Just warn user and process it as much as it can.
*/
if (f_header.data.size == 0) {
fprintf(stderr, "WARNING: The %s file's data size field is 0 which is unexpected.\n"
"Was the 'perf record' command properly terminated?\n",
input_fname);
}
/* Copy the event attribute section */
{
/* SANITY CHECK.
* If these two quantities are not equal, we must
* be processing the perf file format that this tool
* does not support.
*/
if(f_header.attr_size != sizeof(struct perf_file_attr))
{
fprintf(stderr, "header attr_size (%" PRIu64 ") not equal to "
"the size of struct perf_file_attr (%" PRIu64 "). "
"Your perf.data file is the format that this tool "
"does not understand. Sorry!\n",
f_header.attr_size, sizeof(struct perf_file_attr));
exit(-1);
}
/* Now read and copy file sections corresponding to individual attributes.
* We read and copy attributes one by one, because each attribute contains
* a pointer to another file location containing more data. And we need to
* copy that data as well.
*/
int i, nr_attrs = f_header.attrs.size / f_header.attr_size;
for(i = 0; i < nr_attrs; i++)
{
perf_file_attr f_attr;
struct perf_event_attr *attr = &(f_attr.attr);
lseek(ifd, f_header.attrs.offset + i*f_header.attr_size, SEEK_SET);
lseek(ofd, f_header.attrs.offset + i*f_header.attr_size, SEEK_SET);
read_and_exit_on_error(ifd, &f_attr, f_header.attr_size, __FILE__, __LINE__);
write_and_exit_on_error(ofd, &f_attr, f_header.attr_size, __FILE__, __LINE__);
printf("Set to offset %ld and read %ld bytes of perf_file_attr (%ld size)\n",
f_header.attrs.offset + i*f_header.attr_size, f_header.attr_size,
sizeof(f_attr));
if(f_attr.ids.size > 0)
{
printf("There's %" PRId64 " bytes of data at offset %" PRId64 "\n",
f_attr.ids.size,
f_attr.ids.offset);
void *buffer = malloc_and_exit_on_error(f_attr.ids.size, __FILE__, __LINE__);
lseek(ifd, f_attr.ids.offset, SEEK_SET);
lseek(ofd, f_attr.ids.offset, SEEK_SET);
read_and_exit_on_error(ifd, buffer, f_attr.ids.size, __FILE__, __LINE__);
write_and_exit_on_error(ofd, buffer, f_attr.ids.size, __FILE__, __LINE__);
free(buffer);
}
if(!(attr->sample_type & PERF_SAMPLE_TIME) && !attr->sample_id_all)
{
fprintf(stderr, "Event %s does not sample time. "
"We do not know how to process such events.\n",
attr->type < PERF_TYPE_MAX ? event_attr_names[attr->type]: "UNKNOWN");
exit(-1);
}
/* Keep this event attribute in the list. We will later use them to parse samples.
* Technically, we need only the first one (see comment in event_do_we_care()), but
* let's keep them all just in case.
*/
event_descr *attr_to_keep =
malloc_and_exit_on_error(sizeof(event_descr), __FILE__, __LINE__);
attr_to_keep->attr = *attr;
attr_to_keep->sample_size = compute_sample_size(attr->sample_type);
list_insert_and_exit_on_error(&event_attr_list, (void*)attr_to_keep, __FILE__, __LINE__);
printf("Found event %s, sample type is %" PRIu64 ", sample size is %" PRIu64 "\n",
attr->type < PERF_TYPE_MAX ? event_attr_names[attr->type]: "UNKNOWN",
attr->sample_type, attr_to_keep->sample_size);
char *what_we_are_sampling = what_are_we_sampling(attr->sample_type);
printf("%s\n", what_we_are_sampling);
free(what_we_are_sampling);
if(!attr->sample_id_all)
{
fprintf(stderr, "This perf file does not have sample IDs for all data "
"(sample_id_all not set on an event attribute)."
"We rely on sample id timestamp in the COMM event to calibrate "
"timestamps, so this program won't work without sample id data. "
"Try using a more recent version of perf. Sorry!\n");
exit(-1);
}
}
}
/* Copy the event section */
{
void *buffer = malloc_and_exit_on_error(f_header.event_types.size, __FILE__, __LINE__);
lseek(ifd, f_header.event_types.offset, SEEK_SET);
lseek(ofd, f_header.event_types.offset, SEEK_SET);
read_and_exit_on_error(ifd, buffer, f_header.event_types.size, __FILE__, __LINE__);
write_and_exit_on_error(ofd, buffer, f_header.event_types.size, __FILE__, __LINE__);
printf("read event_types: %ld bytes at offset %ld\n", f_header.event_types.size,
f_header.event_types.offset);
free(buffer);
}
/* Cull the data section.
* This section is structured as a collection of perf_event records. Each
* record begins with a header, which has a size field. We don't know in advance how large a
* record is or how it is structured.
* So we are going to read the header first to find out the size and type, and then we read
* the rest.
*/
{
size_t bytes_processed = 0, bytes_written_to_manicured_file = 0;
/* We position the input and output file at the start of the data section,
* but from now on, the files may not be moving synchronously if we are
* copying only selected records to the output file.
*/
lseek(ifd, f_header.data.offset, SEEK_SET);
lseek(ofd, f_header.data.offset, SEEK_SET);
/* We now copy records one by one and decide if we care about them. */
while(bytes_processed < f_header.data.size)
{
union perf_event *event;
perf_event_header *event_header = (perf_event_header*) &event;
size_t this_event_size;
read_and_exit_on_error(ifd, &event, sizeof(perf_event_header), __FILE__, __LINE__);
/* Ok, now we know the size of this event record */
this_event_size = event_header->size;
event = malloc(this_event_size);
/* Read the entire event from the file.
* We don't have to re-read the header, because we already have it,
* but let's do it anyway, this makes the code simpler.
*/
lseek(ifd, -sizeof(perf_event_header), SEEK_CUR);
read_and_exit_on_error(ifd, event, this_event_size, __FILE__, __LINE__);
/* Ok, now determine if we care about this event.
* Its timestamp must fall between the begin and end timestamps
* supplied as arguments.
*/
if(event_do_we_care(event, begin_time, end_time))
{
write_and_exit_on_error(ofd, event, this_event_size, __FILE__, __LINE__);
bytes_written_to_manicured_file += this_event_size;
}
bytes_processed += this_event_size;
printf("IF offset: %ld, OF offset: %ld\n,", lseek(ifd, 0, SEEK_CUR),
lseek(ofd, 0, SEEK_CUR));
free(event);
}
printf("data section: %ld bytes at offset %ld. Processed %ld bytes \n",
f_header.data.size, f_header.data.offset, bytes_processed);
/* Now let's re-write the file header section of the output file
* to update the data section size.
*/
f_header_manicured = f_header;
f_header_manicured.data.size = bytes_written_to_manicured_file;
lseek(ofd, 0, SEEK_SET);
write_and_exit_on_error(ofd, &f_header_manicured, sizeof(perf_file_header), __FILE__, __LINE__);
}
/* Copy the additional features section.
* This section begins at the end of the data section and has a number of
* records of type perf_file_section. How many records there are is determined
* by the number of set bits in the adds_features bitmap in the file header.
*
* If we have not copied the entire data section from the input file to the
* output file, the output file will have a hole in it. That's ok.
*
* These additional features (defined in tools/perf/util/header.h) include
* things like hostname, OS release, NUMA topology, etc. So we copy them
* unchanged into the manicured file.
*/
{
size_t feat_offset = f_header.data.offset + f_header.data.size;
int i, nr_records;
nr_records = bitmap_weight(f_header.adds_features, HEADER_FEAT_BITS);
/* Each record in this section is a perf_file_section, so it
* has a pointer to another file location with data. So we
* process these records one by one and copy the record itself and
* the data to which it points.
*/
for(i = 0; i < nr_records; i++)
{
perf_file_section rec;
/* For the output file we use the same offset as the
* one in the original input file, even though the
* manicured output file is most likely shorter than
* the original (because we skipped data records).
* That is probably okay. Only means that our output file
* will have a "hole" in it.
*/
lseek(ifd, feat_offset + i * sizeof(rec), SEEK_SET);
lseek(ofd, feat_offset + i * sizeof(rec), SEEK_SET);
read_and_exit_on_error(ifd, &rec, sizeof(rec), __FILE__, __LINE__);
write_and_exit_on_error(ofd, &rec, sizeof(rec), __FILE__, __LINE__);
printf("Adds feats: read %ld bytes at offset %ld\n", sizeof(rec),
feat_offset + i * sizeof(rec));
printf("There's %ld more bytes at offset %ld\n", rec.size, rec.offset);
lseek(ifd, rec.offset, SEEK_SET);
lseek(ofd, rec.offset, SEEK_SET);
void *buffer = malloc_and_exit_on_error(rec.size, __FILE__, __LINE__);
read_and_exit_on_error(ifd, buffer, rec.size, __FILE__, __LINE__);
write_and_exit_on_error(ofd, buffer, rec.size, __FILE__, __LINE__);
free(buffer);
}
}
}