Leaks/Status Meetings/2007-October-9

From MozillaWiki
Jump to: navigation, search

« Back to Status Meetings

Agenda

Finding plans

  • Schrep's post in moz.dev.performance [1]
    • Should we do a CoScripter thing?

Tinderbox tests

  • Status for tinderbox tests.
  • Should we force CC between page-loads during perf tests? bug 399107

Status on found leak bugs

  • Status on Peter's cycle collection changes?
  • Are there leak bugs filed that are blocking further discovery?

To Do

  • Get leak monitoring as an extension so that QA can test for leaks during litmus.
    • bug 394466 should fix OS leaks on linux.
    • AI:cbarrett Look into OS leaks on mac.
  • Improve tinderbox tests.
    • Need to measure leaked GDI/user/kernel handles on windows. Possibly show on tinderbox.
    • bug 374822 Need a windows tinderbox with leak numbers since OS leaks are lower there.
    • AI:sicking Do leak monitoring while running tinderbox tests.
      • We do want to run trace-refcnt, trace-malloc on reftest/mochitest. (bug 397724)
      • Ideally we'd shut down between each testcase to identify where leaks happen. Mostly useful until we reach 0, but would help with regressions too.
    • Want a separate test suite that tests for assertions, crashes, leaks. (bug 397725)
  • Important leak bugs:
    • bug 396368 Live connect leak the biggest one still not fixed.
    • bug 398292 tracks bugs that make it difficult to find other leaks through fuzzing.
    • bug 394052 Seems to be the only remaining window/document/docshell leak in Mochitests on at least some platforms, possibly all of them.
  • Finding leaks in extensions
    • Need to solve license issue

Action Items

  • AI:sicking Look into enabling trace-refcount, without the tracing part in nightlies.
  • AI:cbarrett Look into OS leaks on Mac.
  • AI:sicking Talk with build about setting up a Windows leak test machine (bug 374822).

Meeting Notes

  • CoScripter good idea, better for regressions than finding new leaks
  • Force CC during Tp tests
    • Want to measure CC times separately
    • Figure out why smaugs CC patch has effect before turning it off
      • Add logging so we can see what happens on Tp box
    • Easier to see real growth issues if you do CC often
  • Peters patch bug 379718
    • Integrate CC and GC in one operation, rather than GC first and CC after.
    • Makes us not have refcount js objects
    • Fixes most of the issue with having to cyclecollect twice to get all garbage.
      • Could still happen with too weak unlink code.
    • Hopefully ready this week.
  • QA to run litmus with debug builds and look for leaks
    • Checking with build if they can generate the needed binaries
    • Debug builds might be too big for the community. Look into creating special builds without symbols to get down size.