Messages in this thread | | | Date | Thu, 18 Jul 2013 17:27:01 +0200 | From | Oleg Nesterov <> | Subject | Re: [RFC PATCH 0/4] tracing: fix open/delete fixes |
| |
On 07/18, Masami Hiramatsu wrote: > > (2013/07/17 23:43), Oleg Nesterov wrote: > > Once again, I am still not sure and I am asking for your review. > > OK,
Good ;)
> > - If we kill .open/release, we do not need the nontrivial > > refcounting. Everything becomes simple, no need to keep > > the state "in between". > > That also means to refrain checking existence under locking mutex > in all operations.
Speaking of event_enable_write() it needs the same mutex anyway.
> And we have to check it, which I actually concern. > refcounting is not so small and itself is complex, but it just > needs to inc/dec on open/close.
And this inc/dec needs event_mutex too, and the code is not trivial.
But yes, sure, I am not saying that it is always a win performance-wise. In particular, with the patches I sent event/format holds event_mutex between .start and .stop. But again, this is only to make the patch simple. We can narrow the scope of this lock, we can switch to i_mutex (needs the trivial change in invalidate_event_files) which should not be contended.
And of course, sometimes it is better to do the "hard work" in .open() and make .read/write as fast/simple as possible. But not in event/* case, I think.
> > - This also simplifies trace_remove_event_call() paths, we > > know that once it takes event_mutex nobody can play with > > ftrace_event_file/ftrace_event_call we are going to free. > > Hmm, it seems that we can remove only refcount check, or more?
But this check is not necessarily trivial too.
And to remind, personally I do not really like the fact that the opened file blocks rmdir or unregister_probe_event().
To summarise. I believe that this approach is better (and simpler) in general. But I understand that "better" is subjective, so I won't argue. Not to mention, it can be simply wrong so I will heavily rely on your/Steven's review anyway.
Oleg.
| |