Re: [PATCH 1/4] perf tools: Remove /SYSV from no_dso maps

[Joe Mario <jmario@redhat.com>]

On 10/6/25 2:01 PM, Ian Rogers wrote:
> On Mon, Oct 6, 2025 at 10:57 AM Michael Petlan <mpetlan@redhat.com> wrote:
>>
>> We will need it for upcoming commits. Also adjust the function
>> calls in perf-inject, in order not to break the logic.
>>
>> Suggested-by: Jiri Olsa <jolsa@kernel.org>
>>
>> Signed-off-by: Michael Petlan <mpetlan@redhat.com>
>> ---
>>  tools/perf/builtin-inject.c | 4 ++--
>>  tools/perf/util/map.c       | 5 +++--
>>  tools/perf/util/map.h       | 6 +++++-
>>  3 files changed, 10 insertions(+), 5 deletions(-)
>>
>> diff --git a/tools/perf/builtin-inject.c b/tools/perf/builtin-inject.c
>> index a114b3fa1bea..cc090da5c445 100644
>> --- a/tools/perf/builtin-inject.c
>> +++ b/tools/perf/builtin-inject.c
>> @@ -773,7 +773,7 @@ static int tool__inject_build_id(const struct perf_tool *tool,
>>
>>         if (is_anon_memory(filename) || flags & MAP_HUGETLB)
>>                 return 0;
>> -       if (is_no_dso_memory(filename))
>> +       if (is_no_dso_memory(filename) || is_shared_memory(filename))
>>                 return 0;
>>
>>         if (inject->known_build_ids != NULL &&
>> @@ -813,7 +813,7 @@ static int tool__inject_mmap2_build_id(const struct perf_tool *tool,
>>         /* Return to repipe anonymous maps. */
>>         if (is_anon_memory(filename) || flags & MAP_HUGETLB)
>>                 return 1;
>> -       if (is_no_dso_memory(filename))
>> +       if (is_no_dso_memory(filename) || is_shared_memory(filename))
>>                 return 1;
>>
>>         if (dso__read_build_id(dso)) {
>> diff --git a/tools/perf/util/map.c b/tools/perf/util/map.c
>> index b46c68c24d1c..6d13b4602970 100644
>> --- a/tools/perf/util/map.c
>> +++ b/tools/perf/util/map.c
>> @@ -133,12 +133,13 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,
>>         if (ADD_RC_CHK(result, map)) {
>>                 char newfilename[PATH_MAX];
>>                 struct dso *dso;
>> -               int anon, no_dso, vdso, android;
>> +               int anon, no_dso, vdso, android, shared;
>>
>>                 android = is_android_lib(filename);
>>                 anon = is_anon_memory(filename) || flags & MAP_HUGETLB;
>>                 vdso = is_vdso_map(filename);
>>                 no_dso = is_no_dso_memory(filename);
>> +               shared = is_shared_memory(filename);
>>                 nsi = nsinfo__get(thread__nsinfo(thread));
>>
>>                 if ((anon || no_dso) && nsi && (prot & PROT_EXEC)) {
>> @@ -174,7 +175,7 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,
>>                 assert(!dso__kernel(dso));
>>                 map__init(result, start, start + len, pgoff, dso, prot, flags);
>>
>> -               if (anon || no_dso) {
>> +               if (anon || no_dso || shared) {
>>                         map->mapping_type = MAPPING_TYPE__IDENTITY;
>>
>>                         /*
>> diff --git a/tools/perf/util/map.h b/tools/perf/util/map.h
>> index 9cadf533a561..616f32504e46 100644
>> --- a/tools/perf/util/map.h
>> +++ b/tools/perf/util/map.h
>> @@ -258,10 +258,14 @@ static inline int is_anon_memory(const char *filename)
>>  static inline int is_no_dso_memory(const char *filename)
>>  {
>>         return !strncmp(filename, "[stack", 6) ||
>> -              !strncmp(filename, "/SYSV", 5)  ||
>>                !strcmp(filename, "[heap]");
>>  }
>>
>> +static inline int is_shared_memory(const char *filename)
>> +{
>> +       return !strncmp(filename, "/SYSV", 5);
>> +}
> 
> Shared memory could happen because of two mmaps having the same file
> and offset. Is the mapping shared within or across processes? Perhaps
> there is a clearer and more capable API hiding here.

Hi Ian:
We see it happening both within a process and across processes, where the code does something like this:

    shmid = shmget((key_t)0, 8192, IPC_CREAT|0777); 
    shmat_addr1 = shmat(shmid, (char *)0, 0);
    shmat_addr2 = shmat(shmid, (char *)0, 0);

The separate shmat() calls can either be in the same process or different processes.  As long as they use the same shmid and a NULL address, they end up with different virtual addresses pointing to the same location in shared memory.

Then if cacheline contention occurs from accesses to shmat_addr1 and shmat_addr2, perf c2c cannot detect it.

I hope this helps.
Thanks,
Joe

> 
> Thanks,
> Ian
> 
>> +
>>  static inline void map__set_start(struct map *map, u64 start)
>>  {
>>         RC_CHK_ACCESS(map)->start = start;
>> --
>> 2.47.3
>>
>