Re: [PATCH] perf: do not flush maps on COMM for perf report

[Luigi Semenzato]

On Wed, Aug 22, 2012 at 12:28 AM, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
> * Luigi Semenzato <semenzato@xxxxxxxxxxxx> wrote:
>
>> This fixes a long-standing bug caused by the lack of separate
>> COMM and EXEC record types, which makes "perf report" lose
>> track of symbols when a process renames itself.
>>
>> With this fix (suggested by Stephane Eranian), a COMM (rename)
>> no longer flushes the maps, which is the correct behavior.
>> An EXEC also no longer flushes the maps, but this doesn't
>> matter because as new mappings are created (for the executable
>> and the libraries) the old mappings are automatically removed.
>> This is not by accident: the functionality is necessary because
>> DLLs can be explicitly loaded at any time with dlopen(),
>> possibly on top of existing text, so "perf report" handles
>> correctly the clobbering of new mappings on top of old ones.
>>
>> An alternative patch (which I proposed earlier) would be to
>> introduce a separate PERF_RECORD_EXEC type, but it is a much
>> larger change (about 300 lines) and is not necessary.
>
> It would be nice to add that too - we already have FORK/EXIT,
> this seems like a natural extension.

Yes.  Adding PERF_RECORD_EXEC is/would be the right long-term
solution.  But there are two issues.

1. One nice aspect of perf is that perf.data files and "perf report"
are compatible across a large number of versions.  Adding
PERF_RECORD_EXEC breaks compatibility in a somewhat unpleasant manner.
 New perf.data files won't work with old versions of perf and *might*
fail poorly (segv) although this situation is difficult to analyze.

2. Adding a new record type is messy.  It replicates a lot of
boilerplate code, much of it in the kernel, and affects many parts of
the system.  It adds to size, complexity, and likelihood of new bugs.

I would prioritize the "would be nice" category as follows.

1. Improve the handling of unknown record types for future better
backward compatibility.  (Small change.)

2. Refactor/cleanup code to improve readability and robustness.  (Big
change, but can be broken into many smaller pieces.)

3. Add PERF_RECORD_EXEC.

If there is consensus, I might be able to give a shot to 1 and 2
(courtesy of Google).
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/