Hi Roger, Volker,
Glad to see you guys are receptive to this and that it can move forward. Let me
know if you'd like me to help out, for example with reviews or something.
s'marks
On 7/31/15 9:55 AM, Roger Riggs wrote:
Hi Volker,
I agree that adding an Info.commandline() method would be a good way
to make the command line available and be able to describe that it is
OS dependent and may be truncated.
And having it assemble the command and arguments when they are available makes
sense.
As an API addition it will need a clear spec and I can run it through CCC so it
gets
another review and compatibility tests.
Thanks, Roger
On 7/31/2015 5:03 AM, Volker Simonis wrote:
On Fri, Jul 31, 2015 at 2:51 AM, Stuart Marks <stuart.ma...@oracle.com> wrote:
On 7/29/15 11:36 AM, Volker Simonis wrote:
!! ProcessHandleImpl_unix: 709-738: I still disagree with returning
truncated or incomplete
values for the executable or arguments.
Can anyone be a tie-breaker (with a good rational and suggestion for how
an
application can use the data).
As I wrote, I agree to hear other opinions here.
All I want to say that this truncated data is actually what 'ps' is
reporting on Solaris since decades and people seem to be happy with
it. As use case I can imagine logging or monitoring (something like
'top' in Java) where this data will be just good enough.
We could specially mark possibly incomplete data by extending the Info
object with functions like commandIsExact() or argumentsIsPrecise().
But notice that we can not statically preset these values per
platform. For example on Solaris, the 'command()' would return a
precise value for processes with the same uid like the VM but only
inaccurate values for all other processes. The "arguments()" would be
always inaccurate on Solaris/AIX.
It seems like there are cases where either exact or only approximate
information is available. And as you observed, you might get one or the
other on the same platform, depending on the UID. It also might depend on
process state; I believe that some information becomes inaccessible when the
process enters the zombie state.
I don't think we should simply ignore one case or the other, but I also
don't think we should try to cram the different information into the same
API.
The current ProcessHandle.Info api has
Optional<String> command()
Optional<String[]> arguments()
It sounds to me like Roger wants these to contain only exact information.
That seems reasonable, and this probably needs to be specified more clearly
to contrast with what's below.
On Solaris, the psinfo_t struct has char pr_psargs[PRARGSZ] which is a
single string which appears to be the concatenation of the arguments (maybe
including the command name). It's also truncated to fit PRARGSZ. It doesn't
make sense to me to try to return this as a String[], as the zeroth element
of that array, and certainly not parsed out into "words". So maybe instead
we should have a different API that returns an imprecise command line as a
single string:
Optional<String> cmdline()
(Naming bikeshed TBS). The semantics would be that this is the process'
command and arguments concatenated into a single string (thus potentially
losing argument boundaries) and also possibly truncated based on platform
(COUGHsolarisCOUGH) limitations. It's certainly useful for printing out in a
ps, top, or Activity Monitor style application, for informational purposes.
If this were implemented, then on Solaris, command() and arguments() would
always return empty optionals.
I'm not sure what this should be if the exact information is available. It
would be inconvenient if something that just wanted to print out an
approximate command line had to check several different APIs to get the
information. Maybe cmdline() could assemble the information from exact data
if it's is available, by concatenating the Strings from command() and
arguments(), as a convenience to the caller. But I could go either way on
this.
Not sure this counts as a tie-breaker, but it might be a reasonable way
forward.
s'marks
Hi Stuart,
thanks a lot for your comments - I like your proposal. For me this
sounds like a good compromise.
@Roger: should I go and add a new field commandLine and the
corresponding getter to the Info class? As Stuart proposed, the getter
could check if 'command' and 'arguments' are available and assemble
the command line from them. Otherwise it could use the content of the
commandLine field if that is available.
Regards,
Volker
