Hello,

I've found a couple of bugs in the DWARF location handling and had
some updates to improve the quality of the type resolution.

The current code only checks the first variable it found in the
closest scope but sometimes it's not good because macro expansions
like container_of (which is used by many list/rb tree manipulation
routines) only gives a very limited information (like void *) with
type cast.  So it needs to lookup other variables in the upper scope.
See the patch 8 for more details.

Also sometimes it can have more information for the parent type if the
pointer is for an embedded type.  For example, a list_head is
typically a part of bigger struct.  Even if it found a variable for
the list_head, it'd be nice if it can tell which list it is.

To compare two type information in general, I've added a heuristic to
take a pointer to a bigger data type.

This is an example data, the portion of unknown type went down a bit
and the atomic_t turned out to be _mapcount in the struct page.

Before:
  #
  # Overhead  Data Type
  # ........  .........
  #
      37.24%  (unknown)
      14.40%  atomic_t 
       8.81%  (stack operation)
       5.54%  struct psi_group_cpu
       3.40%  struct task_struct
       2.99%  struct pcpu_hot
       2.99%  struct cfs_rq
       2.18%  struct audit_krule
       1.93%  struct psi_group
       1.62%  struct sched_entity

After:
  #
  # Overhead  Data Type
  # ........  .........
  #
      36.87%  (unknown)
      14.40%  struct page
       8.81%  (stack operation)
       6.00%  struct psi_group_cpu
       3.40%  struct task_struct
       3.36%  struct cfs_rq
       2.99%  struct pcpu_hot
       2.18%  struct audit_krule
       1.93%  struct psi_group
       1.62%  struct sched_entity

Also updated the debug message and the statistics to help debugging.

The code is available at 'perf/data-profile-update-v1' branch in
git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git


Thanks,
Namhyung


Namhyung Kim (9):
  perf dwarf-aux: Check allowed location expressions when collecting
    variables
  perf annotate-data: Fix off-by-one in location range check
  perf annotate-data: Add enum type_match_result
  perf annotate-data: Add variable_state_str()
  perf annotate-data: Change return type of find_data_type_block()
  perf annotate-data: Add is_pointer_type() helper
  perf annotate-data: Add is_better_type() helper
  perf annotate-data: Check variables in every scope
  perf annotate-data: Update type stat at the end of
    find_data_type_die()

 tools/perf/util/annotate-data.c | 359 ++++++++++++++++++++------------
 tools/perf/util/dwarf-aux.c     |   5 +-
 2 files changed, 230 insertions(+), 134 deletions(-)

-- 
2.46.0.184.g6999bdac58-goog

On Fri, Aug 16, 2024 at 04:58:30PM -0700, Namhyung Kim wrote:
> Hello,
> 
> I've found a couple of bugs in the DWARF location handling and had
> some updates to improve the quality of the type resolution.
> 
> The current code only checks the first variable it found in the
> closest scope but sometimes it's not good because macro expansions
> like container_of (which is used by many list/rb tree manipulation
> routines) only gives a very limited information (like void *) with
> type cast.  So it needs to lookup other variables in the upper scope.
> See the patch 8 for more details.
> 
> Also sometimes it can have more information for the parent type if the
> pointer is for an embedded type.  For example, a list_head is
> typically a part of bigger struct.  Even if it found a variable for
> the list_head, it'd be nice if it can tell which list it is.
> 
> To compare two type information in general, I've added a heuristic to
> take a pointer to a bigger data type.
> 
> This is an example data, the portion of unknown type went down a bit
> and the atomic_t turned out to be _mapcount in the struct page.
> 
> Before:
>   #
>   # Overhead  Data Type
>   # ........  .........
>   #
>       37.24%  (unknown)
>       14.40%  atomic_t 
>        8.81%  (stack operation)
>        5.54%  struct psi_group_cpu
>        3.40%  struct task_struct
>        2.99%  struct pcpu_hot
>        2.99%  struct cfs_rq
>        2.18%  struct audit_krule
>        1.93%  struct psi_group
>        1.62%  struct sched_entity
> 
> After:
>   #
>   # Overhead  Data Type
>   # ........  .........
>   #
>       36.87%  (unknown)
>       14.40%  struct page
>        8.81%  (stack operation)
>        6.00%  struct psi_group_cpu
>        3.40%  struct task_struct
>        3.36%  struct cfs_rq
>        2.99%  struct pcpu_hot
>        2.18%  struct audit_krule
>        1.93%  struct psi_group
>        1.62%  struct sched_entity
> 
> Also updated the debug message and the statistics to help debugging.
> 
> The code is available at 'perf/data-profile-update-v1' branch in
> git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
> 

Thanks, applied to perf-tools-next,

- Arnaldo