2018-05-07 18:45:13

by Josh Hunt

[permalink] [raw]
Subject: [PATCH v2] perf tools: allow map files to specify DSO

Add the ability to specify a DSO in the /tmp/perf-<PID>.map file.
The DSO should be the first line in the file and readable by the
running user. If a valid DSO is found all other contents of the
file will be ignored. This allows things like callchain unwinding
with DWARF to work.

Suggested-by: Wang Nan <[email protected]>
Signed-off-by: Josh Hunt <[email protected]>
---
We have an application which uses huge pages for its text section, but
still needs the ability to do callchain unwinding with DWARF. We use
the perf-<PID>.map file setup to do symbol resolution and that works
great, but callchain unwinding fails.

A few months ago I mentioned this to Wang Nan and he suggested a way
around this problem could be to specify the path of the DSO in the map
file. The attached patch is my initial hack at this. Running with this
patch I can now get full callchain unwinding with DWARF.

FWIW LBR + map file works with callchains, but unfortunately there are
some cases where we still need DWARF.

I submitted an RFC as an earlier draft:
https://lkml.org/lkml/2018/4/24/1070

v2:
1. Rebase RFC patch to
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git perf/core
2. Update jit-interface.txt as per Arnaldo's request

tools/perf/Documentation/jit-interface.txt | 8 +++++---
tools/perf/util/map.c | 32 ++++++++++++++++++++++++++++++
2 files changed, 37 insertions(+), 3 deletions(-)

diff --git a/tools/perf/Documentation/jit-interface.txt b/tools/perf/Documentation/jit-interface.txt
index a8656f564915..8a25b979802f 100644
--- a/tools/perf/Documentation/jit-interface.txt
+++ b/tools/perf/Documentation/jit-interface.txt
@@ -3,13 +3,15 @@ by a JIT.

The JIT has to write a /tmp/perf-%d.map (%d = pid of process) file

-This is a text file.
+This is a text file and can have one of the following formats:

-Each line has the following format, fields separated with spaces:
+1) Each line has the following format, fields separated with spaces:

START SIZE symbolname

START and SIZE are hex numbers without 0x.
symbolname is the rest of the line, so it could contain special characters.

-The ownership of the file has to match the process.
+2) A single line with the full pathname of the DSO to use.
+
+For both formats, the ownership of the file has to match the process.
diff --git a/tools/perf/util/map.c b/tools/perf/util/map.c
index c8fe836e4c3c..da3050b18e34 100644
--- a/tools/perf/util/map.c
+++ b/tools/perf/util/map.c
@@ -139,6 +139,31 @@ void map__init(struct map *map, u64 start, u64 end, u64 pgoff, struct dso *dso)
refcount_set(&map->refcnt, 1);
}

+static bool replace_anon(char *mapfilename)
+{
+ FILE *file = NULL;
+ bool ret = false;
+
+ file = fopen(mapfilename, "r");
+ if (file != NULL) {
+ char *line = NULL;
+ size_t line_len, linesz=0;
+
+ line_len = getline(&line, &linesz, file);
+ if (line_len > 0) {
+ line[line_len-1] = '\0'; /* null terminate */
+ if (!access(line, R_OK)) {
+ strlcpy(mapfilename, line, line_len);
+ ret = true;
+ }
+ }
+ free(line);
+ fclose(file);
+ }
+
+ return ret;
+}
+
struct map *map__new(struct machine *machine, u64 start, u64 len,
u64 pgoff, u32 d_maj, u32 d_min, u64 ino,
u64 ino_gen, u32 prot, u32 flags, char *filename,
@@ -170,6 +195,13 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,
snprintf(newfilename, sizeof(newfilename),
"/tmp/perf-%d.map", nsi->pid);
filename = newfilename;
+ /*
+ * Check to see if map file references DSO to use, if so, use it.
+ */
+ if (anon && replace_anon(newfilename)) {
+ anon = 0;
+ filename = newfilename;
+ }
}

if (android) {
--
1.9.1



2018-05-07 18:40:46

by Andi Kleen

[permalink] [raw]
Subject: Re: [PATCH v2] perf tools: allow map files to specify DSO

On Mon, May 07, 2018 at 02:24:16PM -0400, Josh Hunt wrote:
> Add the ability to specify a DSO in the /tmp/perf-<PID>.map file.
> The DSO should be the first line in the file and readable by the
> running user. If a valid DSO is found all other contents of the
> file will be ignored. This allows things like callchain unwinding
> with DWARF to work.

FWIW it's ok, but also obsolete with Kirill's large-pages-in-tmpfs
work in newer kernels. With that you can just copy the executable into
a 2MB tmpfs and disable the manual huge page copying and everything
should work as usually.

So essentially it's only a hack for old kernels and old binaries.

But doesn't hurt I guess.

-Andi

2018-05-07 21:36:09

by Josh Hunt

[permalink] [raw]
Subject: Re: [PATCH v2] perf tools: allow map files to specify DSO

On 05/07/2018 11:40 AM, Andi Kleen wrote:
> On Mon, May 07, 2018 at 02:24:16PM -0400, Josh Hunt wrote:
>> Add the ability to specify a DSO in the /tmp/perf-<PID>.map file.
>> The DSO should be the first line in the file and readable by the
>> running user. If a valid DSO is found all other contents of the
>> file will be ignored. This allows things like callchain unwinding
>> with DWARF to work.
>
> FWIW it's ok, but also obsolete with Kirill's large-pages-in-tmpfs
> work in newer kernels. With that you can just copy the executable into
> a 2MB tmpfs and disable the manual huge page copying and everything
> should work as usually.
>
> So essentially it's only a hack for old kernels and old binaries.
>
> But doesn't hurt I guess.

Ah, very interesting. I wasn't aware of this. Can you point me to some
more details on this process?

Thanks
Josh

2018-05-08 00:35:59

by Andi Kleen

[permalink] [raw]
Subject: Re: [PATCH v2] perf tools: allow map files to specify DSO

On Mon, May 07, 2018 at 02:30:32PM -0700, Josh Hunt wrote:
> On 05/07/2018 11:40 AM, Andi Kleen wrote:
> > On Mon, May 07, 2018 at 02:24:16PM -0400, Josh Hunt wrote:
> > > Add the ability to specify a DSO in the /tmp/perf-<PID>.map file.
> > > The DSO should be the first line in the file and readable by the
> > > running user. If a valid DSO is found all other contents of the
> > > file will be ignored. This allows things like callchain unwinding
> > > with DWARF to work.
> >
> > FWIW it's ok, but also obsolete with Kirill's large-pages-in-tmpfs
> > work in newer kernels. With that you can just copy the executable into
> > a 2MB tmpfs and disable the manual huge page copying and everything
> > should work as usually.
> >
> > So essentially it's only a hack for old kernels and old binaries.
> >
> > But doesn't hurt I guess.
>
> Ah, very interesting. I wasn't aware of this. Can you point me to some more
> details on this process?

See commit 5a6e75f8110c97e2a5488894d4e922187e6cb343

-Andi