Re: [PATCH v5 13/13] perf record: introduce --ctl-fd[-ack] options

[Jiri Olsa <jolsa@redhat.com>]

On Fri, Jun 05, 2020 at 05:47:28PM +0300, Alexey Budankov wrote:
> 
> On 05.06.2020 16:57, Jiri Olsa wrote:
> > On Fri, Jun 05, 2020 at 04:15:52PM +0300, Alexey Budankov wrote:
> >>
> >> On 05.06.2020 13:51, Jiri Olsa wrote:
> >>> On Tue, Jun 02, 2020 at 04:43:58PM +0300, Adrian Hunter wrote:
> >>>> On 2/06/20 12:12 pm, Alexey Budankov wrote:
> >>>>>
> >>>>> On 02.06.2020 11:32, Alexey Budankov wrote:
> >>>>>>
> >>>>>> On 02.06.2020 2:37, Andi Kleen wrote:
> >>>>>>>>> or a pathname, or including also the event default of "disabled".
> >>>>>>>>
> >>>>>>>> For my cases conversion of pathnames into open fds belongs to external
> >>>>>>>> controlling process e.g. like in the examples provided in the patch set.
> >>>>>>>> Not sure about "event default of 'disabled'"
> >>>>>>>
> >>>>>>> It would be nicer for manual use cases if perf supported the path names
> >>>>>>> directly like in Adrian's example, not needing a complex wrapper script.
> >>>>>>
> >>>>>> fds interface is required for VTune integration since VTune wants control
> >>>>>> over files creation aside of Perf tool process. The script demonstrates
> >>>>>> just one possible use case.
> >>>>>>
> >>>>>> Control files could easily be implemented on top of fds making open operations
> >>>>>> for paths and then initializing fds. Interface below is vague and with explicit
> >>>>>> options like below it could be more explicit:
> >>>>>> --ctl-file /tmp/my-perf.fifo --ctl-file-ack /tmp/my-perf-ack.fifo
> >>>>>
> >>>>> Or even clearer:
> >>>>>
> >>>>> --ctl-fifo /tmp/my-perf --ctl-fifo-ack /tmp/my-perf-ack
> >>>>
> >>>> If people are OK with having so many options, then that is fine by me.
> >>>
> >>> the single option Adrian suggested seems better to me:
> >>>
> >>>  --control
> >>>  --control 11
> >>>  --control 11,15
> >>
> >> What if a user specifies fifos named like this above, not fds?
> >>
> >>>  --control 11,15,disabled
> >>>  --control 11,,disabled
> >>>  --control /tmp/my-perf.fifo
> >>>  --control /tmp/my-perf.fifo,/tmp/my-perf-ack.fifo
> >>
> >> What if a user wants not fifos but other type of comm channels?
> >>
> >>>  --control /tmp/my-perf.fifo,/tmp/my-perf-ack.fifo,disabled
> >>>  --control /tmp/my-perf.fifo,,disabled
> >>>
> >>> we already support this kind of options arguments, like for --call-graph
> >>>
> >>> jirka
> >>>
> >>
> >> IMHO,
> >> this interface, of course, looks more compact (in amount of options) however
> >> the other side it is less user friendly. One simple option for one simple
> >> purpose is more convenient as for users as for developers. Also complex
> >> option syntax tends to have limitations and there are probably more
> >> non-obvious ones.
> >>
> >> Please speak up. I might have missed something meaningful.
> > 
> > how about specify the type like:
> > 
> > --control fd:1,2,...
> 
> What do these ... mean?

other possible options

> 
> > --control fifo:/tmp/fifo1,/tmp/fifo2
> > --control xxx:....
> > 
> > this way we can extend the functionality in the future
> > and stay backward compatible, while keeping single option
> 
> Well, it clarifies more. However it still implicitly assumes
> and requires proper ordering e.g. 1 is ctl-fd and 2 is ack-fd
> and if there are some more positions there will be gaps like
> --control fd:10,,something,,something ...

right, that's what we do for other options

> 
> Why is one single option with complex syntax more preferable
> than several simple options? Also it would still consume almost
> equal amount of command line space in shell.

I think it's better for future.. say if there's going to be support
for passing file paths you'll need to add something like --ctl-fifo
and --ctl-fifo-ack no?  with single option we'd just add something
like:

  --control fifo:/tmp/my-perf.fifo,/tmp/my-perf-ack.fifo

jirka