RFR (M/L): 8131168: Refactor ProcessHandleImpl_*.c and add implememtation for AIX
Volker Simonis
volker.simonis at gmail.com
Mon Aug 10 17:13:36 UTC 2015
On Fri, Aug 7, 2015 at 8:54 PM, Roger Riggs <Roger.Riggs at oracle.com> wrote:
> Hi Volker,
>
> 1) ProcessHandle.java:243
>
> For the definition of the new commandLine method I would:
>
> - Use @link instead of @code for references to commands() and arguments()
> for easy navigation
>
> - @implNote[1] should I think be changed to @apiNote:
> the text describes not the JDK implementation but is information about the
> information returned and is use to the application developer, not the JDK
> implementation.
>
> - The specific references to Linux implementation command line length
> parameters seems out of place
> and should be omitted.
>
> /**
> * Returns the command line of the process.
> * <p>
> * If {@link #command command()} and {@link #arguments arguments()}
> return non-null
> * optionals, this is simply a convenience method which concatenates
> * the values of the two functions separated by spaces. Otherwise it
> will return a
> * best-effort, platform dependent representation of the command
> line.
> *
> * @apiNote Note that the returned executable pathname and the
> * arguments may be truncated on some platforms due to
> system
> * limitations.
> * <p>
> * The executable pathname may contain only the
> * name of the executable without the full path
> information.
> * It is undecideable whether white space separates
> different
> * arguments or is part of a single argument.
> *
> * @return an {@code Optional<String>} of the command line
> * of the process
> */
> public Optional<String> commandLine();
>
>
> ProcessHandle.java:252: in arguments() method - @apiNote is a better fit for
> the note
>
> ProcessHandleImpl_macosx.c:276: - indentation +4
>
Thanks for the correction. I've taken your wording you suggested.
>
> 2) ProcessHandleImpl_macosx.c:192:
> if (errno != EINVAL) ... There was previously this test, I'm concerned
> that if the pid is invalid,
> it will now throw a RuntimeException instead of returning -1.
> I recall a discussion from May that recommended testing for EINVAL.
> The sysctl in getCmdlineAndUserInfo also does not throw if errno !=
> EINVAL, so the usage
> is not consistent (probably my coding) but needs investigation.
>
Not sure about this one and couldn't find any previous discussion
about the topic.
But, according to the sysctl man-page, EINVAL is only returned if:
- The name array is less than two or greater than CTL_MAXNAME.
- A non-null newp is given and its specified length in newlen is too
large or too small.
The first case can not happen because we always statically allocate
arrays of the correct size.
The second case can not happen as well, because we always have 'newp == NULL'.
So according to this information I don't see any reason why we should
check for EINVAL. I think the right solution is to check for 'oldlenp
> 0' which we already do. By the way, this is also the check applied
by the psutils (see the implementation of 'get_kinfo_proc()' in [1]).
So I wnated to also removed the last check for EINVAL in
getCmdlineAndUserInfo(). But for some reason, that seems to be really
necessary. Without it, we will get a RuntinmeException if we call
sysinfo for pid==0 for example. Further research showed that the
kernel seems to really return EINVAL for KERN_PROCARGS2 (see function
sysctl_procargsx() in [2]). But KERN_PROCARGS2 isn't specified as a
supported constant in the sysctl man-page, so the information there is
still valid :)
On the other hand, I found that the psutils alo handles EINVAL only
for KERN_PROCARGS2 (see get_arg_list() in [1]).
So to cut a long story short, I think the current implementation is
safe as it is now!
[1] http://psutil.googlecode.com/svn/trunk/psutil/arch/osx/process_info.c
[2] http://www.opensource.apple.com/source/xnu/xnu-1699.24.23/bsd/kern/kern_sysctl.c
> 3) ProcessHandleImpl_solaris.c can do without the includes:
> #include "jni_util.h"
> #include "java_lang_ProcessHandleImpl.h"
> #include "java_lang_ProcessHandleImpl_Info.h"
>
> #include <stdio.h>
>
> 4) Ditto ProcessHandleImpl_aix.c
>
Thanks, fixed.
> 5) ProcessHandleImpl_unix.c: 618: typo "fuctions" -> "functions"
>
Fixed.
> 6) ProcessHandleImpl_unix.c:463: Would it be worthwhile to put
> sysconf(_SC_GETPW_R_SIZE_MAX) in a static?
>
Good point! While doing this I realized that 'clock_ticks_per_second'
is only used on Linux. So I moved the declaration of
'clock_ticks_per_second' from ProcessHandleImpl_unix.hpp to
ProcessHandleImpl_linux.cpp and its initialization to os_initNative()
in the same file.
I also declare a new static 'getpw_buf_size' in
ProcessHandleImpl_unix.cpp and initialize it in
Java_java_lang_ProcessHandleImpl_initNative() in the same file.
> 7) OnExitTest.java: exits without an error, just output in the log; it would
> escape attention.
> The code respects the timeout setting.
> Suggest removing the 'return' @133; the test will produce some errors
> and when debugging
> the note will be in the log.
>
Done.
> 8) ProcessHandleImpl_unix.h:58 - update the comment to include 0 as a valid
> parent pid.
>
Fixed.
> Looks pretty good,
> Thanks, Roger
>
Thanks:)
I've also added two rudimentary tests for the new commandLine()
function to InfoTest.test2():
- if both, 'command()' and 'arguments()' are available, I check that
'commandLine()' starts with 'command()' and contains all the arguments
from 'arguments()'.
- otherwise, if 'commandLine()' is available, I check that it at
least contains 'java'. And as long as it is big enough it should also
contain the corresponding arguments.
The two new tests have been verified to pass on Windows, Linux, MacOS
X, Solaris and AIX.
The new version can be found here:
http://cr.openjdk.java.net/~simonis/webrevs/2015/8131168.v6/
Regards,
Volker
>
>
> [1] http://openjdk.java.net/jeps/8068562 - JEP draft: javadoc tags to
> distinguish API, implementation, specification, and notes
>
>
> On 8/5/2015 3:56 PM, Volker Simonis wrote:
>
> Hi,
>
> so here's the webrev which implements the new Info.commandLine()
> method (I chose 'commandLine() ' instead of 'cmdline()' or
> 'commandline()' because the other getters are camel case as well):
>
> http://cr.openjdk.java.net/~simonis/webrevs/2015/8131168.v4/
>
> From the JavaDoc of the new method:
>
> * If {@code command()} and {@code arguments()} return non-null
> * optionals, this is simply a convenience method which concatenates
> * the values of the two functions. Otherwise it will return a
> * best-effort, platform dependent representation of the command line.
>
> This didn't change anything on MacOS X where there is no additional
> effort to get the command line.
>
> On Solaris and AIX, Info.commandLine() will always return the contents
> of psinfo.pr_psargs because there's no other method to get the exact
> arguments (i.e. Info.arguments() always returns NULL. I could
> therefore remove the extra handling of AIX/Solaris in the InfoTest
> from my initial change.
>
> On Linux, things are a little more complicated:
>
> - the initial implementation for Linux used arg[0] as 'command' if
> /proc/pid/exe wasn't readable. This was true for all the processes we
> don't own (except if we are running as root). But this information may
> be incomplete, because arg[0] may only contain the command without its
> full path. One example is 'sendmail' for which Info.command() reported
> "sendmail: accepting connections" but Info.arguments() was empty. This
> is wrong, because 'sendmail' changes its argv[] array. I have
> therefore disabled this behavior now that we have the 'commandLine()'
> method.
>
> - /proc/pid/cmdline is limited to PAGE_SIZE (normally 4096) characters
> on Linux. So strictly speaking, this isn't 'exact' information as well
> (there are plenty of complains that especially for Java programs this
> is not enough) and should go to 'commandLine()' instead to 'arguments'
> if /proc/pid/cmdline is truncated. I do check for this now.
>
> - the information in /proc/pid/cmdline can also be changed to
> something other than the original arguments if a program changes
> argv[] (which is not forbidden) but there's probably not much we can
> do to detect this. I've added a corresponding @implNote comment to
> JavaDoc of Info.arguments().
>
> - the initial implementation did not check for incomplete reads of
> /proc/pid/cmdline. This may be a problem on systems with PAGE_SIZE >
> 4096 (on Linux/ppc64 a page size of 65536 is not unusual). I'm now
> always reading the complete contents of /proc/pid/cmdline.
>
> - as far as I understand the current implementation, 'arguments()'
> returns the arguments array WITHOUT 'arg[0]' (which is the program
> name) but may
> be we should specify that more clearly in the JavaDoc of 'arguments()'.
>
> That's it. Hope you like it :)
>
> Regards,
> Volker
>
> On Fri, Jul 31, 2015 at 11:02 PM, Stuart Marks <stuart.marks at oracle.com>
> wrote:
>
> 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.marks at 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
>
>
More information about the core-libs-dev
mailing list