Messages in this thread | | | Date | Fri, 11 Dec 2015 09:53:36 +0800 | From | "Wangnan (F)" <> | Subject | Re: [PATCH perf/core 00/22] perf refcnt debugger API and fixes |
| |
On 2015/12/10 23:12, 'Arnaldo Carvalho de Melo' wrote:
[SNIP] > But this requires having these special refcnt__ routines, that will make > tools/perf/ code patterns for reference counts look different that the > refcount patterns in the kernel :-\ > > And would be a requirement to change the observed workload :-\ > > Is this _strictly_ required?
No. The requirement should be:
1. The create/get/put/delete functions are non-inline (because dwarf info is not as reliable as symbol); 2. From their argument list, we can always get the variable we need (the pointer of objects, the value of refcnt, etc.)
We don't have to use this refcnt things.
> Can't we, for a project like perf, where we > know where some refcount (say, the one for 'struct thread') gets > initialized, use that (thread__new()) and then hook into thread__get and > thread__put and then use the destructor, thread__delete() as the place > to dump leaks? >
I think it is possible. If we can abstract a common pattern about it, we can provide a perf subcommand which we can deal with generic alloc/free pattern. I'll put it on my todo-list. Currently we are focusing on perf daemonization.
Thank you.
| |