Skip to contentSkip to conversation

perf: fix race in build_id_cache__add_s()

PATCH Applied Milos Vyletel 4 participants 4 messages Mar 20, 2015 – Mar 22, 2015 perf events

lore
Share & tools
lore

Milos Vyletel

author
#0 lore raw

PATCH 1/1 perf: fix race in build_id_cache__add_s()

int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
const char *name, bool is_kallsyms, bool is_vdso)
{
...

        if (access(filename, F_OK)) {
               ^--------------------------------------------------------- [1]
                if (is_kallsyms) {
                         if (copyfile("/proc/kallsyms", filename))
                                goto out_free;
                } else if (link(realname, filename) && copyfile(name, filename))
                             ^-----------------------------^------------- [2]
                                                            \------------ [3]
                        goto out_free;
        }

...

when multiple instances of perf record get to [1] at more or less same time and run access() one or more may get failure because the file does not exist yet (since the first instance did not have chance to link it yet). at this point the race moves to link() at [2] where first thread to get there links file and goes on but second one gets -EEXIST so it runs copyfile [3] which truncates the file.

recreator:
rm -rf /root/.debug
for cpu in $(awk '/processor/ {print $3}' /proc/cpuinfo); do

	perf record -a -v -T -F 1000 -C $cpu \
		-o perf-${cpu}.data sleep 5 2> /dev/null &

done
wait

and simply search for empty files by
find /lib/modules/uname -r/kernel/* -size 0

Signed-off-by Milos Vyletel <[email protected]>
---
 tools/perf/util/build-id.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/util/build-id.c +2 -1
index 0c72680..a9db5a1 100644
@@ -293,7 +293,8 @@ int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
 		if (is_kallsyms) {
 			 if (copyfile("/proc/kallsyms", filename))
 				goto out_free;
-		} else if (link(realname, filename) && copyfile(name, filename))
+		} else if (link(realname, filename) && errno != EEXIST &&
+				copyfile(name, filename))
 			goto out_free;
 	}
Signature · 1 line
2.1.0
Acked-by on v1 1/1
#1 lore raw
Quoted 14 lines · Milos Vyletel

int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
const char *name, bool is_kallsyms, bool is_vdso)
{
...

        if (access(filename, F_OK)) {
               ^--------------------------------------------------------- [1]
                if (is_kallsyms) {
                         if (copyfile("/proc/kallsyms", filename))
                                goto out_free;
                } else if (link(realname, filename) && copyfile(name, filename))
                             ^-----------------------------^------------- [2]
                                                            \------------ [3]
                        goto out_free;
        }
Quoted 7 lines

...

when multiple instances of perf record get to [1] at more or less same time and run access() one or more may get failure because the file does not exist yet (since the first instance did not have chance to link it yet). at this point the race moves to link() at [2] where first thread to get there links file and goes on but second one gets -EEXIST so it runs copyfile [3] which truncates the file.

nice.. :-\

Acked-by Jiri Olsa <[email protected]>

in addition we should use some inter-perf lock
on all .debug dir operations

jirka

maintainer Applied
#2 lore raw

Em Fri, Mar 20, 2015 at 01:18:07PM +0100, Jiri Olsa escreveu:

Quoted 23 lines · Milos Vyletel
Quoted 14 lines

int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
const char *name, bool is_kallsyms, bool is_vdso)
{
...

        if (access(filename, F_OK)) {
               ^--------------------------------------------------------- [1]
                if (is_kallsyms) {
                         if (copyfile("/proc/kallsyms", filename))
                                goto out_free;
                } else if (link(realname, filename) && copyfile(name, filename))
                             ^-----------------------------^------------- [2]
                                                            \------------ [3]
                        goto out_free;
        }
Quoted 7 lines

...

when multiple instances of perf record get to [1] at more or less same time and run access() one or more may get failure because the file does not exist yet (since the first instance did not have chance to link it yet). at this point the race moves to link() at [2] where first thread to get there links file and goes on but second one gets -EEXIST so it runs copyfile [3] which truncates the file.

Quoted 6 lines

nice.. :-\

Acked-by Jiri Olsa <[email protected]>

in addition we should use some inter-perf lock
on all .debug dir operations

Yeah, would be nice to have that improved.

Thanks, applied.

- Arnaldo

merge bot
#3 lore raw

[tip:perf/core] perf tools: Fix race in build_id_cache__add_s()

In reply to Milos Vyletel
perf tools: Fix race in build_id_cache__add_s()
Commit-ID:  0635b0f71424be7706793ac260d063491a2889a0
Gitweb:     http://git.kernel.org/tip/0635b0f71424be7706793ac260d063491a2889a0
Author:     Milos Vyletel <[email protected]>
AuthorDate: Fri, 20 Mar 2015 11:37:25 +0100
Committer:  Arnaldo Carvalho de Melo <[email protected]>
CommitDate: Fri, 20 Mar 2015 17:49:50 -0300

perf tools: Fix race in build_id_cache__add_s()

int build_id_cache__add_s(const char *sbuild_id, const char *debugdir,
const char *name, bool is_kallsyms, bool is_vdso)
{
...

        if (access(filename, F_OK)) {
               ^--------------------------------------------------------- [1]
                if (is_kallsyms) {
                         if (copyfile("/proc/kallsyms", filename))
                                goto out_free;
                } else if (link(realname, filename) && copyfile(name, filename))
                             ^-----------------------------^------------- [2]
                                                            \------------ [3]
                        goto out_free;
        }

...

When multiple instances of perf record get to [1] at more or less same time and run access() one or more may get failure because the file does not exist yet (since the first instance did not have chance to link it yet).

At this point the race moves to link() at [2] where first thread to get there links file and goes on but second one gets -EEXIST so it runs copyfile [3] which truncates the file.

reproducer:

rm -rf /root/.debug
for cpu in $(awk '/processor/ {print $3}' /proc/cpuinfo); do

	perf record -a -v -T -F 1000 -C $cpu \
		-o perf-${cpu}.data sleep 5 2> /dev/null &

done
wait

and simply search for empty files by:

find /lib/modules/uname -r/kernel/* -size 0

Signed-off-by Milos Vyletel <[email protected]>
Acked-by Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Stephane Eranian <[email protected]>
Link http://lkml.kernel.org/r/[email protected] open here
Signed-off-by Arnaldo Carvalho de Melo <[email protected]>
---
 tools/perf/util/build-id.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/util/build-id.c +2 -1
index a196746..f7fb258 100644
@@ -374,7 +374,8 @@ int build_id_cache__add_s(const char *sbuild_id, const char *name,
 		if (is_kallsyms) {
 			 if (copyfile("/proc/kallsyms", filename))
 				goto out_free;
-		} else if (link(realname, filename) && copyfile(name, filename))
+		} else if (link(realname, filename) && errno != EEXIST &&
+				copyfile(name, filename))
 			goto out_free;
 	}