lkml.org 
[lkml]   [2013]   [Mar]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
SubjectRe: [PATCH] perf: add callgrind conversion tool
Date
Hi Roberto,

On Wed, 27 Feb 2013 00:26:33 +0100, Roberto Vitillo wrote:
> The proposed patch adds the convert tool to perf which allows to convert a
> perf.data file to a callgrind data file which can subsequently be displayed
> with kcachegrind. Callgraphs are not supported.

I was thinking about a similar tool but a little bit more generic way -
it might support more formats (ctf, pprof, trace-cmd?) in both direction
(convert from/to) - without considering its feasibility ;-)

Anyway, nice work!

>
> In order to speed up the converter I am not piping to addr2line, this
> means that
> the code depends on libbfd. Considering that perf doesn't strictly depend on
> libbfd my question is if it would be feasible to add that requirement.
>
> Note that the code may trigger the following bug in libbfd:
> http://sourceware.org/bugzilla/show_bug.cgi?id=15106

Oh, I have to admit first that I don't know about the callgrind, bfd and
binutils. Please feel free to ignore me if I go to wrong direction.

>
> Signed-off-by: Roberto Agostino Vitillo <ravitillo@lbl.gov>
> ---
[SNIP]
> diff --git a/tools/perf/Makefile b/tools/perf/Makefile
> index a2108ca..55c35d5 100644
> --- a/tools/perf/Makefile
> +++ b/tools/perf/Makefile
> @@ -399,6 +399,8 @@ LIB_H += util/intlist.h
> LIB_H += util/perf_regs.h
> LIB_H += util/unwind.h
> LIB_H += util/vdso.h
> +LIB_H += util/cgcnv.h
> +LIB_H += util/a2l.h
> LIB_H += ui/helpline.h
> LIB_H += ui/progress.h
> LIB_H += ui/util.h
> @@ -472,6 +474,8 @@ LIB_OBJS += $(OUTPUT)util/rblist.o
> LIB_OBJS += $(OUTPUT)util/intlist.o
> LIB_OBJS += $(OUTPUT)util/vdso.o
> LIB_OBJS += $(OUTPUT)util/stat.o
> +LIB_OBJS += $(OUTPUT)util/cgcnv.o
> +LIB_OBJS += $(OUTPUT)util/a2l.o
>
> LIB_OBJS += $(OUTPUT)ui/setup.o
> LIB_OBJS += $(OUTPUT)ui/helpline.o
> @@ -528,6 +532,7 @@ BUILTIN_OBJS += $(OUTPUT)builtin-kmem.o
> BUILTIN_OBJS += $(OUTPUT)builtin-lock.o
> BUILTIN_OBJS += $(OUTPUT)builtin-kvm.o
> BUILTIN_OBJS += $(OUTPUT)builtin-inject.o
> +BUILTIN_OBJS += $(OUTPUT)builtin-convert.o

You can make these conditional after checking availibility of bfd.


> BUILTIN_OBJS += $(OUTPUT)tests/builtin-test.o
>
> PERFLIBS = $(LIB_FILE) $(LIBTRACEEVENT)
> @@ -793,39 +798,35 @@ endif
>
> ifdef NO_DEMANGLE
> BASIC_CFLAGS += -DNO_DEMANGLE
> -else
> - ifdef HAVE_CPLUS_DEMANGLE
> +endif
> +
> +FLAGS_BFD=$(ALL_CFLAGS) $(ALL_LDFLAGS) $(EXTLIBS) -DPACKAGE='perf' -lbfd -ldl

Shouldn't it be "perf" rather than 'perf'? I know it doesn't matter
much but looks wierd to me. Please see below:

https://lkml.org/lkml/2012/9/25/453


> +has_bfd := $(call try-cc,$(SOURCE_BFD),$(FLAGS_BFD))
> +ifeq ($(has_bfd),y)
> + EXTLIBS += -lbfd -ldl
> +
> + ifdef HAVE_CPLUS_DEMANGLE
> EXTLIBS += -liberty
> - BASIC_CFLAGS += -DHAVE_CPLUS_DEMANGLE
> - else
> - FLAGS_BFD=$(ALL_CFLAGS) $(ALL_LDFLAGS) $(EXTLIBS) -DPACKAGE='perf' -lbfd
> - has_bfd := $(call try-cc,$(SOURCE_BFD),$(FLAGS_BFD),libbfd)
> - ifeq ($(has_bfd),y)
> - EXTLIBS += -lbfd
> + endif
> +else
> + FLAGS_BFD_IBERTY=$(FLAGS_BFD) -liberty
> + has_bfd_iberty := $(call try-cc,$(SOURCE_BFD),$(FLAGS_BFD_IBERTY))
> +
> + ifeq ($(has_bfd_iberty),y)
> + EXTLIBS += -lbfd -ldl -liberty
> + else
> + FLAGS_BFD_IBERTY_Z=$(FLAGS_BFD_IBERTY) -lz
> + has_bfd_iberty_z := $(call try-cc,$(SOURCE_BFD),$(FLAGS_BFD_IBERTY_Z))
> + ifeq ($(has_bfd_iberty_z),y)
> + EXTLIBS += -lbfd -ldl -liberty -lz
> else
> - FLAGS_BFD_IBERTY=$(FLAGS_BFD) -liberty
> - has_bfd_iberty := $(call try-cc,$(SOURCE_BFD),$(FLAGS_BFD_IBERTY),liberty)
> - ifeq ($(has_bfd_iberty),y)
> - EXTLIBS += -lbfd -liberty
> - else
> - FLAGS_BFD_IBERTY_Z=$(FLAGS_BFD_IBERTY) -lz
> - has_bfd_iberty_z := $(call try-cc,$(SOURCE_BFD),$(FLAGS_BFD_IBERTY_Z),libz)
> - ifeq ($(has_bfd_iberty_z),y)
> - EXTLIBS += -lbfd -liberty -lz
> - else
> - FLAGS_CPLUS_DEMANGLE=$(ALL_CFLAGS) $(ALL_LDFLAGS) $(EXTLIBS) -liberty
> - has_cplus_demangle := $(call
> try-cc,$(SOURCE_CPLUS_DEMANGLE),$(FLAGS_CPLUS_DEMANGLE),demangle)
> - ifeq ($(has_cplus_demangle),y)
> - EXTLIBS += -liberty
> - BASIC_CFLAGS += -DHAVE_CPLUS_DEMANGLE
> - else
> - msg := $(warning No bfd.h/libbfd found, install
> binutils-dev[el]/zlib-static to gain symbol demangling)
> - BASIC_CFLAGS += -DNO_DEMANGLE
> - endif
> - endif
> - endif
> + msg := $(error No bfd.h/libbfd found, install binutils-dev[el]/zlib-static)

It should be a warning and proceed with disabling 'convert' tool.


> endif
> endif
> +
> + ifndef NO_DEMANGLE
> + BASIC_CFLAGS += -DHAVE_CPLUS_DEMANGLE
> + endif
> endif
>
> ifeq ($(NO_PERF_REGS),0)
> diff --git a/tools/perf/builtin-convert.c b/tools/perf/builtin-convert.c
> new file mode 100644
> index 0000000..9311b3a
> --- /dev/null
> +++ b/tools/perf/builtin-convert.c
> @@ -0,0 +1,217 @@
> +/*
> + * builtin-convert.c
> + *
> + * Builtin convert command: Convert a perf.data input file
> + * to a callgrind profile data file.
> + */
[SNIP]
> +static int process_sample_event(struct perf_tool *tool,
> + union perf_event *event,
> + struct perf_sample *sample,
> + struct perf_evsel *evsel,
> + struct machine *machine)
> +{
> + struct perf_convert *cnv = container_of(tool, struct perf_convert, tool);
> + struct addr_location al;
> +
> + if (perf_event__preprocess_sample(event, machine, &al, sample,
> + symbol__annotate_init) < 0) {
> + pr_warning("problem processing %d event, skipping it.\n",
> + event->header.type);
> + return -1;
> + }
> +
> + if (cnv->cpu_list && !test_bit(sample->cpu, cnv->cpu_bitmap))
> + return 0;
> +
> + if (!al.filtered && cg_cnv_sample(evsel, sample, &al)) {

AFAICS this cg_cnv_sample() does nothing with converting. Why did you
move the code to a different file rather than keeping it together?


> + pr_warning("problem incrementing symbol count, skipping event\n");
> + return -1;
> + }
> +
> + return 0;
> +}
> +
> +static void hists__find_annotations(struct hists *self, int evidx,
> + struct perf_convert *cnv)

The name of the function doesn't look good to me. Maybe hists__convert_symbols?


> +{
> + struct rb_node *nd = rb_first(&self->entries);
> + int key = K_RIGHT;

For the purpose of this tool, I guess using input/collapsed tree (i.e.
->entries_in or ->entries_collpased) instead of output tree (->entries)
is more appropriate since it'll collect related dsos and symbols together.


> +
> + while (nd) {
> + struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
> + struct annotation *notes;
> +
> + if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned) {
> + cg_cnv_unresolved(cnv->output_file, evidx, he);
> + goto find_next;
> + }
> +
> + notes = symbol__annotation(he->ms.sym);
> +
> + if (notes->src == NULL) {
> +find_next:
> + if (key == K_LEFT)
> + nd = rb_prev(nd);
> + else
> + nd = rb_next(nd);
> + continue;
> + }
> +
> + cg_cnv_symbol(cnv->output_file, he->ms.sym, he->ms.map);
> +
> + free(notes->src);
> + notes->src = NULL;
> + }
> +}

The traversal of the hist_entry looks strange. How about this?

while (nd) {
struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
struct annotation *notes;

nd = rb_next(nd);

if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned) {
cg_cnv_unresolved(cnv->output_file, evidx, he);
continue;
}

notes = symbol__annotation(he->ms.sym);
if (notes->src == NULL)
continue;

cg_cnv_symbol(cnv->output_file, he->ms.sym, he->ms.map);

free(notes->src);
notes->src = NULL;
}


[SNIP]
> diff --git a/tools/perf/util/a2l.c b/tools/perf/util/a2l.c
> new file mode 100644
> index 0000000..82f1023
> --- /dev/null
> +++ b/tools/perf/util/a2l.c
> @@ -0,0 +1,136 @@
> +/* Based on addr2line. */
> +
> +#include "a2l.h"
> +
> +#define PACKAGE 'perf'
Please see my comment above.

> +
> +#include <bfd.h>
> +#include <stdlib.h>
> +#include <stdio.h>
> +
> +#include <linux/kernel.h>
> +
> +static const char *filename;
> +static const char *functionname;
> +static unsigned int line;
> +static asymbol **syms;
> +static bfd_vma pc;
> +static bfd_boolean found;
> +static bfd *abfd;
[SNIP]
> +static void find_address_in_section(bfd *self, asection *section,
> + void *data ATTRIBUTE_UNUSED)
> +{
> + bfd_vma vma;
> + bfd_size_type size;
> +
> + if (found)
> + return;
> +
> + if ((bfd_get_section_flags(abfd, section) & SEC_ALLOC) == 0)
> + return;

In this function, abfd == self, right? If so, please keep using just
one of them for consistency.

> +
> + vma = bfd_get_section_vma(abfd, section);

Ditto.

> + if (pc < vma)
> + return;
> +
> + size = bfd_get_section_size(section);
> + if (pc >= vma + size)
> + return;
> +
> + found = bfd_find_nearest_line(self, section, syms, pc - vma,

Ditto.

> + &filename, &functionname, &line);
> +}
> +
> +int addr2line_init(const char *file_name)
> +{
> + abfd = bfd_openr(file_name, NULL);
> + if (abfd == NULL)
> + return -1;
> +
> + if (!bfd_check_format(abfd, bfd_object))
> + return bfd_fatal(bfd_get_filename(abfd));
> +
> + return slurp_symtab();
> +
> +}
> +
> +void addr2line_cleanup(void)
> +{
> + if (syms != NULL) {
> + free(syms);
> + syms = NULL;
> + }
> +
> + bfd_close(abfd);
> + line = found = 0;
> +}
> +
> +int addr2line_inline(const char **file, unsigned *line_nr)
> +{
> +
> + found = bfd_find_inliner_info(abfd, &filename, &functionname, &line);
> + *file = filename;
> + *line_nr = line;

Why not using 'file' and 'line_nr' directly?

> +
> + return found;
> +}
> +
> +int addr2line(unsigned long addr, const char **file, unsigned *line_nr)
> +{
> + found = 0;
> + pc = addr;
> + bfd_map_over_sections(abfd, find_address_in_section, NULL);
> +
> + *file = filename;
> + *line_nr = line;
> +
> + return found;
> +}
[SNIP]
> diff --git a/tools/perf/util/cgcnv.c b/tools/perf/util/cgcnv.c
> new file mode 100644
> index 0000000..02fcfa5
> --- /dev/null
> +++ b/tools/perf/util/cgcnv.c
> @@ -0,0 +1,188 @@
> +#include "cgcnv.h"
> +#include "a2l.h"
> +#include "parse-events.h"
> +#include "symbol.h"
> +#include "map.h"
> +#include "util.h"
> +#include "annotate.h"
> +
> +#include <linux/list.h>
> +#include <linux/rbtree.h>
> +
> +static const char *last_source_name;
> +static unsigned last_line;
> +static u64 last_off;
> +
> +int cg_cnv_header(FILE *output, struct perf_session *session)
> +{
> + struct perf_evsel *pos;
> +
> + fprintf(output, "positions: instr line\nevents:");
> + list_for_each_entry(pos, &session->evlist->entries, node) {
> + const char *evname = NULL;
> + struct hists *hists = &pos->hists;
> + u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
> +
> + if (nr_samples > 0) {
> + evname = perf_evsel__name(pos);
> + fprintf(output, " %s", evname);
> + }

This will skip events with no samples, but later codes simply output
all events.


> + }
> + fprintf(output, "\n");
> +
> + return 0;
> +}
> +
> +int cg_cnv_sample(struct perf_evsel *evsel, struct perf_sample *sample,
> + struct addr_location *al)
> +{
> + struct hist_entry *he;
> + int ret = 0;
> +
> + he = __hists__add_entry(&evsel->hists, al, NULL, 1);
> + if (he == NULL)
> + return -ENOMEM;
> +
> + ret = 0;
> + if (he->ms.sym != NULL) {
> + struct annotation *notes = symbol__annotation(he->ms.sym);
> + if (notes->src == NULL && symbol__alloc_hist(he->ms.sym) < 0)
> + return -ENOMEM;
> +
> + ret = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
> + }
> +
> + evsel->hists.stats.total_period += sample->period;
> + hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
> + return ret;
> +}
> +
> +static void cg_sym_header_printf(FILE *output, struct symbol *sym,
> + struct map *map, struct annotation *notes,
> + u64 offset)
> +{
> + int idx, ret, ret_callee, ret_caller = 0;
> + u64 address = map__rip_2objdump(map, sym->start) + offset;
> + unsigned caller_line;
> + const char *caller_name;
> +
> + ret_callee = addr2line(address, &last_source_name, &last_line);
> + while ((ret = addr2line_inline(&caller_name, &caller_line)))
> + ret_caller = ret;
> +
> + /* Needed to display correctly the inlining relationship in kcachegrind */
> + if (ret_caller && caller_line)
> + fprintf(output, "fl=%s\n0 0\n", caller_name);
> +
> + if (ret_callee && last_line)
> + fprintf(output, "fl=%s\n", last_source_name);
> + else
> + fprintf(output, "fl=\n");

Could you explain why this empty fl line is needed?

> +
> + fprintf(output, "%#" PRIx64 " %u", address, last_line);
> + for (idx = 0; idx < notes->src->nr_histograms; idx++)
> + fprintf(output, " %" PRIu64,
> + annotation__histogram(notes, idx)->addr[offset]);
> +
> + fprintf(output, "\n");
> + last_off = offset;
> +}
> +
> +static void cg_sym_events_printf(FILE *output, struct symbol *sym,
> + struct map *map, struct annotation *notes,
> + u64 offset)
> +{
> + int ret, idx;
> + unsigned line;
> + const char *filename;
> +
> + ret = addr2line(map__rip_2objdump(map, sym->start) + offset,
> + &filename, &line);
> + if (filename && last_source_name && strcmp(filename, last_source_name)) {
> + fprintf(output, "fl=%s\n", filename);
> + last_source_name = filename;

In this case the 'ret' is always 0? Otherwise, I'm not sure the 'last_line'
still has valid line number.

Thanks,
Namhyung


> + }
> +
> + if (ret)
> + fprintf(output, "+%" PRIu64 " %+d", offset - last_off,
> + (int)(line - last_line));
> + else
> + fprintf(output, "+%" PRIu64 " %u", offset - last_off, line);
> +
> + for (idx = 0; idx < notes->src->nr_histograms; idx++) {
> + u64 cnt = annotation__histogram(notes, idx)->addr[offset];
> + fprintf(output, " %" PRIu64, cnt);
> + }
> +
> + fprintf(output, "\n");
> + last_off = offset;
> + last_line = line;
> +}


\
 
 \ /
  Last update: 2013-03-04 04:21    [W:0.067 / U:1.200 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site