RFR(S): 8241638: launcher time metrics alway report 1 on Linux when _JAVA_LAUNCHER_DEBUG set(Internet mail)

[David Holmes]

On 2/04/2020 6:48 pm, linzang(臧琳) wrote:
> Hi David,
>      Thanks to point it out, I have uploaded a new patch at http://cr.openjdk.java.net/~lzang/8241638/webrev03/

As Alan pointed out this definition is not just for Linux. I would 
suggest this comment block:

!  * Add CounterGet() implementation for launch time debug on Linux.
!  * Use gettimeofday() here since the gethrtime() or clock_gettime()
!  * may not be available for all Linux platforms.
!  * The potential risk of using gettimeofday() is it can be affected
!  * by settimeofday() or adjtime().
!  * Choose gettimeofday() here because it is more common on linux and
!  * it is better than just return magic numbers for launch time debug.

becomes simply:

* Provide a CounterGet() implementation based on gettimeofday() which
* is universally available, even though it may not be 'high resolution'
* compared to platforms that provide gethrtime() (like Solaris). It is
* also subject to time-of-day changes, but alternatives may not be
* known to be available at either build time or run time.

No need for updated webrev.

Thanks,
David

>   
> BRs,
> Lin
> 
> On 2020/4/2, 2:54 PM, "David Holmes" <david.holmes at oracle.com> wrote:
> 
>      Hi Lin,
>      
>      On 31/03/2020 12:39 pm, linzang(臧琳) wrote:
>      > Hi David, Henry, Alan,
>      >      Thanks a lot for your review.
>      >      I have considered use clock_gettime() first, but I seached the code and found there is a marco SUPPORTS_CLOCK_MONOTONIC guard the usage of it. Which makes me think it may not be available under specific situation. So I choosed gettimeofday.
>      >     Do you think the patch need to be refined to remove HAVE_GETHRTIME as Alan suggested?  Thanks.
>      
>      Leaving it as-is seems okay. Though the likelihood of anyone taking
>      advantage of an existing gethrtime() function anywhere other than
>      Solaris seems near zero. But if the Solaris port deprecation goes ahead
>      the point will be moot as only the "linux" version will remain.
>      
>      This comment block needs fixing up:
>      
>        37 /*
>        38  *  * Add gethrtime() implementation for launch time debug on Linux.
>        39  *   */
>      
>      Also a nit but the new function can be called whatever you like as you
>      are defining CounterGet, so calling it gethrtime() is a bit misleading -
>      I suggest getTimeMicros().
>      
>      Thanks,
>      David
>      
>      > BRs,
>      > Lin
>      >
>      > >On 2020/3/31, 8:05 AM, "David Holmes" <david.holmes at oracle.com> wrote:
>      >>
>      >>     On 31/03/2020 4:08 am, Henry Jen wrote:
>      >>     > Based on my understanding to gethrtime(), the main benefit is not to be affected by settimeofday or adjtime. I think it is probably better to use
>      >>     >
>      >>     > clock_gettime(CLOCK_MONOTONIC_RAW, ts);
>      >>     >
>      >>     > which I checked seems to be available on both Linux and Mac. Haven’t test it though.
>      >>
>      >>     Not guaranteed to be available - either clock_gettime function or that
>      >>     particular clock - at build time or runtime. We use a check in the build
>      >>     system to determine build-time availability for hotspot, and then use
>      >>     dl_lookup etc at runtime to determine if actually available. We should
>      >>     be able to get rid of this one day but we checked fairly recently and
>      >>     there were still some issues.
>      >
>      >>     gettimeofday is a lot better than returning 1. Otherwise call into the
>      >>     VM and use JVM_NanoTime.
>      >>
>      >>     Cheers,
>      >>     David
>      >>     -----
>      >>
>      >>     > Cheers,
>      >>     > Henry
>      >>     >
>      >>     >> On Mar 30, 2020, at 1:37 AM, Alan Bateman <Alan.Bateman at oracle.com> wrote:
>      >>     >>
>      >>     >> On 30/03/2020 03:41, linzang(臧琳) wrote:
>      >>     >>> Dear All,
>      >>     >>>       May I ask your help to reivew this tiny patch? Thanks.
>      >>     >>>
>      >>     >>>
>      >>     >>>
>      >>     >>> BRs,
>      >>     >>> Lin
>      >>     >>>
>      >>     >>> From: "linzang(臧琳)" <linzang at tencent.com>
>      >>     >>> Date: Thursday, March 26, 2020 at 3:13 PM
>      >>     >>> To: "core-libs-dev at openjdk.java.net" <core-libs-dev at openjdk.java.net>
>      >>     >>> Subject: RFR(S): 8241638: launcher time metrics alway report 1 on Linux when _JAVA_LAUNCHER_DEBUG set
>      >>     >>>
>      >>     >>> Dear All,
>      >>     >>> May I ask your help to review this tiny fix?
>      >>     >>>      Bug: https://bugs.openjdk.java.net/browse/JDK-8241638
>      >>     >>>      Webrev: http://cr.openjdk.java.net/~lzang/8241638/webrev01/
>      >>     >>> Thanks!
>      >>     >>>
>      >>     >> Using gettimeofday on non-Solaris platforms seems reasonable here. The comment in the patch suggests Linux but it's other Unix builds too. Also just a minor nit that the code in java.base uses 4-space indent, not 2. Looking at the patch makes me wondering if we should remove HAVE_GETHRTIME as it seems to be only used on Solaris >and the launcher is already using #ifdef __solaris__ in several places. Henry, do you have any comments on this?
>      >   >   >>
>      >   >   >> -Alan
>      >   >   >
>      >
>      >
>      >
>      
>      
>