Arnaldo,
A long time ago, I brought up the issue of perf failing to resolve
data symbols despite having the MAP__VARIABLE map type.
Now that I have posted the PEBS-LL patch series, there is a test
case to debug this.
I looked again at it today. Start from ip__resolve_data() in my patchset.
IT is looking for a data symbol given an address.
Seems to me the core of the problem is how perf processes the MMAP
records in machine__process_mmap_event():
int machine__process_mmap_event(struct machine *machine, union
perf_event *event)
{
...
thread = machine__findnew_thread(machine, event->mmap.pid);
if (thread == NULL)
goto out_problem;
map = map__new(&machine->user_dsos, event->mmap.start,
event->mmap.len, event->mmap.pgoff,
event->mmap.pid, event->mmap.filename,
MAP__FUNCTION);
I don't see any effort to distinguish code from data here. All mmaps are
treated as code.
That could be fine, if then you drop MAP__VARIABLE.
So how is this supposed to work?
Thanks.
On Tue, Oct 30, 2012 at 9:29 PM, Stephane Eranian <[email protected]> wrote:
> Arnaldo,
>
> A long time ago, I brought up the issue of perf failing to resolve
> data symbols despite having the MAP__VARIABLE map type.
>
> Now that I have posted the PEBS-LL patch series, there is a test
> case to debug this.
>
> I looked again at it today. Start from ip__resolve_data() in my patchset.
> IT is looking for a data symbol given an address.
>
> Seems to me the core of the problem is how perf processes the MMAP
> records in machine__process_mmap_event():
>
> int machine__process_mmap_event(struct machine *machine, union
> perf_event *event)
> {
> ...
>
> thread = machine__findnew_thread(machine, event->mmap.pid);
> if (thread == NULL)
> goto out_problem;
> map = map__new(&machine->user_dsos, event->mmap.start,
> event->mmap.len, event->mmap.pgoff,
> event->mmap.pid, event->mmap.filename,
> MAP__FUNCTION);
>
> I don't see any effort to distinguish code from data here. All mmaps are
> treated as code.
>
> That could be fine, if then you drop MAP__VARIABLE.
>
> So how is this supposed to work?
>
Looking at the kernel code for attr->mmap_data shows a corollary
problem with the RECORD_MMAP. Somehow we lose the fact
that this corresponds to data mmap (!VM_EXEC). Having that
info in the header would help the tool process the record properly,
I think.