115690 - clean up tinderbox script for win32 perf tests (gmake build)

Status: RESOLVED FIXED

(Webtools Graveyard :: Tinderbox, defect)

Description

[Cathleen]

We need to start working on getting perf tests hooked up to windows machines.
mcafee said that tinderbox script for windows has bit-rotted...   :-(  so, it
needs work.  He will need to do a re-write, in order to get it into the shape
linux script is.  

(linux script has a lot of unix specific stuff, therefore it is not sharable)

Comment 1

[Chris McAfee]

cls's win32 gmake work might be a good solution here
future, off 098 radar

Comment 2

[Chris McAfee]

requesting cls to post patch here, so jrgm & us can help get Ts test going.

Comment 3

[Chris McAfee]

-> seawood

Comment 4

[hacker formerly known as [e-mail]]

Attachment: Initial hack to get Alive & Bloat tests running with win32 gmake build

This is the original hack that I wacked into place Monday night and has been
running for 2 days now.  It still needs to be cleaned up as I hardcoded some
win32 paths (like location of the profile).

Comment 5

[Suresh Duddi (gone)]

This is great. Mcafee tells me timer tests hangs. Could I help.

Jrgm, we need to install HiRes timer module right ?

Comment 6

[hacker formerly known as [e-mail]]

Well, it looks like AliveTestReturnToken() isn't exiting correctly after the
first run.  The app doesn't exit.  In the second run, a second window will popup
and in subsequent runs, the test exits with a "failed" status.  I'm noticing
something similar on the ports boxes.  BeOS has occassionally had problems
killing off BloatTests when the load got too high but we usually don't have
problems with Solaris. Yet, I just killed off about 20 mozilla-bin
file:/blah/../startup.html?23423423" runs from nebiros.

Is there any special build config that's needed to run the startup tests?  

Comment 7

[Chris McAfee]

No special requirements, you need to pull
mozilla/tools/performance/startup/startup-test.html separately,
then just turn the test on.  The tbox script should turn on the 
JS dump pref automatically so we can scrape for the numbers.

Comment 8

[hacker formerly known as [e-mail]]

Attachment: win32 tests v1.3

I cleaned up the patch a bit.  I replaced the hardcoded values with
Settings::OS checks where possible.  If anyone has been paying attention to the
Ports page, you'll see that we actually had BloatTest2 running at one point. 
It only seems to run for 2 cycles at the most before it starts failing.

Comment 9

[hacker formerly known as [e-mail]]

Attachment: win32 tests v1.4

The only change from the previous patch is that it doesn't add --shutdown-leaks
to the BloatTest2 cmdline if $Settings::OS WINNT.  W/o --shutdown-leaks,
BloatTest2 runs fine in multiple succession.

Comment 10

[John Morrison]

Attachment: incremental diffs to cls' patch v1.4

I extended cls' patch to work on win98 (yes, perverse, but I would like to be
able to run continuous tests on win98, although I hope to make it so that I
don't have to actually build on win98, but just have the script take delivery
of a build from a (faster) winnt tinderbox).

Anyways, the changes make it so that the 'eq 'WINNT'' tests are now '=~
/^WIN/',
properly dealing with the uname output of newer cygwin, some allowances for
command.com braindeadness, setting the errorparser/buildfamily to 'win32', and 

allowing the user to define $Settings:Timezone, because POSIX::strftime says 
"Pacific Standard Time" and my version of cvs insists on "PST".

I probably have some more work to do w.r.t. correctly finding the win98 profile
directory. I also need to work on using Win32::Process to run the AliveTest,
etc., since 'fork and exec' just does not want to work on good old DOS98.

I've tested the overall patch (to be attached) on Linux, and there are no
problems there (and of course it runs correctly on win98 modulo the fact that 
win98 goes nuts after about 10 consecutive builds).

Comment 11

[John Morrison]

Attachment: cls v1.4 patch plus Attachment #73169

cls' patch with my changes also added.

Comment 12

[John Morrison]

checked in attachment 73170 [details] [diff] [review], r=mcafee.

Comment 13

[cls]

-> jrgm

Comment 14

[Brian Ryner (not reading)]

I figured out the problem with this today.  The startup time test URL contains
shell metacharacters, so perl 'exec' uses /bin/sh to execute the command.  When
this happens, for some reason, the child process of /bin/sh can't be killed (I
also can't kill it using cygwin's "kill" command).  Instead, we can put the
arguments into an array ourself, which will prevent perl from using the shell to
run the command.  I'll post a patch after I get it cleaned up.

Comment 15

[Brian Ryner (not reading)]

Attachment: patch v1

With this patch I'm able to successfully run Txul and Ts on Windows with Cygwin
perl and the HiRes perl module installed.

Comment 16

[Brian Ryner (not reading)]

Attachment: patch to build-ns post-build script

since I'm too lazy to file a separate bug for this...

Comment 17

[Brian Ryner (not reading)]

Attachment: updated patch

This patch does the following:

- converts all commands to an arguments array for passing to exec()
- Runs "dos2unix" on the log file before mailing it off, if on Windows
- Adds a "ProductName" config option to control where to look for the profile
on Windows.
- Fixes clobbering of the profile on Windows.
- Makes the script look for the HiRes time module instead of the Win32API
module for cygwin perl.

Comment 18

[hacker formerly known as [e-mail]]

Comment on attachment 95957 [details] [diff] [review]
updated patch

With the exception of the the @@ -1568,7 +1581,7 @@ chunk of
build-seamonkey-util.pl, r=cls.  $hostname isn't defined anywhere.

Comment 19

[hacker formerly known as [e-mail]]

Attachment 95957 [details] [diff] has been checked in.

Comment 20

[cls]

perf tests have been running on win32 for some time now.