The addr2line command is started and then addresses piped to it. In
order to determine the end of a addr2lines output a ',' it output with
an expectation to get '??\n??:0\n' as a reply. llvm-addr2line differs
in that ',' generates a reply of ','.
The approach detects and then caches the addr2line style. When records
are read the sentinel is detected appropriately.
Comparing the output there is a little more inline data on my machine
with llvm-addr2line:
$ sudo perf record -a -g sleep 1
$ sudo perf report --addr2line=addr2line > a.txt
$ sudo perf report --addr2line=llvm-addr2line > b.txt
$ wc -l a.txt b.txt
12386 a.txt
12477 b.txt
Some other small changes, switching to the api/io code to avoid file
streams wrapping the command's stdin/stdout. Ignore SIGPIPE for when
addr2line exits and writes fail.
v2. Address review comments from Arnaldo and Namhyung, fixing a
realloc error path, argument ordering and a comment.
Ian Rogers (4):
tools api: Add io__getline
perf srcline: Simplify addr2line subprocess
perf srcline: Support for llvm-addr2line
perf srcline: Avoid addr2line SIGPIPEs
tools/lib/api/io.h | 45 ++++++++++
tools/perf/tests/api-io.c | 36 ++++++++
tools/perf/util/srcline.c | 171 +++++++++++++++++++++++---------------
3 files changed, 184 insertions(+), 68 deletions(-)
--
2.40.0.348.gf938b09366-goog
Ignore SIGPIPEs when addr2line is configured.
Signed-off-by: Ian Rogers <[email protected]>
---
tools/perf/util/srcline.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/tools/perf/util/srcline.c b/tools/perf/util/srcline.c
index f4fcdada821b..cfca03abd6f8 100644
--- a/tools/perf/util/srcline.c
+++ b/tools/perf/util/srcline.c
@@ -473,6 +473,8 @@ static enum a2l_style addr2line_configure(struct child_process *a2l)
ch = io__get_char(&io);
} while (ch > 0 && ch != '\n');
}
+ /* Ignore SIGPIPE in the event addr2line exits. */
+ signal(SIGPIPE, SIG_IGN);
}
return style;
}
--
2.40.0.348.gf938b09366-goog
Em Mon, Apr 03, 2023 at 11:40:29AM -0700, Ian Rogers escreveu:
> The addr2line command is started and then addresses piped to it. In
> order to determine the end of a addr2lines output a ',' it output with
> an expectation to get '??\n??:0\n' as a reply. llvm-addr2line differs
> in that ',' generates a reply of ','.
>
> The approach detects and then caches the addr2line style. When records
> are read the sentinel is detected appropriately.
>
> Comparing the output there is a little more inline data on my machine
> with llvm-addr2line:
> $ sudo perf record -a -g sleep 1
> $ sudo perf report --addr2line=addr2line > a.txt
> $ sudo perf report --addr2line=llvm-addr2line > b.txt
> $ wc -l a.txt b.txt
> 12386 a.txt
> 12477 b.txt
>
> Some other small changes, switching to the api/io code to avoid file
> streams wrapping the command's stdin/stdout. Ignore SIGPIPE for when
> addr2line exits and writes fail.
>
> v2. Address review comments from Arnaldo and Namhyung, fixing a
> realloc error path, argument ordering and a comment.
Added to local repo, build testing, will be in tmp.perf-tools-next soon.
- Arnaldo
> Ian Rogers (4):
> tools api: Add io__getline
> perf srcline: Simplify addr2line subprocess
> perf srcline: Support for llvm-addr2line
> perf srcline: Avoid addr2line SIGPIPEs
>
> tools/lib/api/io.h | 45 ++++++++++
> tools/perf/tests/api-io.c | 36 ++++++++
> tools/perf/util/srcline.c | 171 +++++++++++++++++++++++---------------
> 3 files changed, 184 insertions(+), 68 deletions(-)
>
> --
> 2.40.0.348.gf938b09366-goog
>
--
- Arnaldo
Hi,
On Mon, Apr 3, 2023 at 1:24 PM Arnaldo Carvalho de Melo <[email protected]> wrote:
>
> Em Mon, Apr 03, 2023 at 11:40:29AM -0700, Ian Rogers escreveu:
> > The addr2line command is started and then addresses piped to it. In
> > order to determine the end of a addr2lines output a ',' it output with
> > an expectation to get '??\n??:0\n' as a reply. llvm-addr2line differs
> > in that ',' generates a reply of ','.
> >
> > The approach detects and then caches the addr2line style. When records
> > are read the sentinel is detected appropriately.
> >
> > Comparing the output there is a little more inline data on my machine
> > with llvm-addr2line:
> > $ sudo perf record -a -g sleep 1
> > $ sudo perf report --addr2line=addr2line > a.txt
> > $ sudo perf report --addr2line=llvm-addr2line > b.txt
> > $ wc -l a.txt b.txt
> > 12386 a.txt
> > 12477 b.txt
> >
> > Some other small changes, switching to the api/io code to avoid file
> > streams wrapping the command's stdin/stdout. Ignore SIGPIPE for when
> > addr2line exits and writes fail.
> >
> > v2. Address review comments from Arnaldo and Namhyung, fixing a
> > realloc error path, argument ordering and a comment.
>
> Added to local repo, build testing, will be in tmp.perf-tools-next soon.
>
> - Arnaldo
>
> > Ian Rogers (4):
> > tools api: Add io__getline
> > perf srcline: Simplify addr2line subprocess
> > perf srcline: Support for llvm-addr2line
> > perf srcline: Avoid addr2line SIGPIPEs
Acked-by: Namhyung Kim <[email protected]>
Thanks,
Namhyung
> >
> > tools/lib/api/io.h | 45 ++++++++++
> > tools/perf/tests/api-io.c | 36 ++++++++
> > tools/perf/util/srcline.c | 171 +++++++++++++++++++++++---------------
> > 3 files changed, 184 insertions(+), 68 deletions(-)
> >
> > --
> > 2.40.0.348.gf938b09366-goog
> >
>
> --
>
> - Arnaldo
Em Mon, Apr 03, 2023 at 10:18:47PM -0700, Namhyung Kim escreveu:
> Hi,
>
> On Mon, Apr 3, 2023 at 1:24 PM Arnaldo Carvalho de Melo <[email protected]> wrote:
> >
> > Em Mon, Apr 03, 2023 at 11:40:29AM -0700, Ian Rogers escreveu:
> > > The addr2line command is started and then addresses piped to it. In
> > > order to determine the end of a addr2lines output a ',' it output with
> > > an expectation to get '??\n??:0\n' as a reply. llvm-addr2line differs
> > > in that ',' generates a reply of ','.
> > >
> > > The approach detects and then caches the addr2line style. When records
> > > are read the sentinel is detected appropriately.
> > >
> > > Comparing the output there is a little more inline data on my machine
> > > with llvm-addr2line:
> > > $ sudo perf record -a -g sleep 1
> > > $ sudo perf report --addr2line=addr2line > a.txt
> > > $ sudo perf report --addr2line=llvm-addr2line > b.txt
> > > $ wc -l a.txt b.txt
> > > 12386 a.txt
> > > 12477 b.txt
> > >
> > > Some other small changes, switching to the api/io code to avoid file
> > > streams wrapping the command's stdin/stdout. Ignore SIGPIPE for when
> > > addr2line exits and writes fail.
> > >
> > > v2. Address review comments from Arnaldo and Namhyung, fixing a
> > > realloc error path, argument ordering and a comment.
> >
> > Added to local repo, build testing, will be in tmp.perf-tools-next soon.
> >
> > - Arnaldo
> >
> > > Ian Rogers (4):
> > > tools api: Add io__getline
> > > perf srcline: Simplify addr2line subprocess
> > > perf srcline: Support for llvm-addr2line
> > > perf srcline: Avoid addr2line SIGPIPEs
>
> Acked-by: Namhyung Kim <[email protected]>
Thanks, added to those patches.
- Arnaldo