Review Request 120420: [OS X] a more complete approach to prevent the crash after finishing a code-difference review or git/commit

Sun Dec 21 23:02:35 UTC 2014

> On Dec. 21, 2014, 10:41 p.m., Milian Wolff wrote:
> > shell/sessioncontroller.cpp, line 323
> > <https://git.reviewboard.kde.org/r/120420/diff/4/?file=333844#file333844line323>
> >
> >     under what circumstances would you get here without a valid document? since the arg is not a smart pointer, a queued connection would still see a "non-zero" document pointer, even when the document got destroyed. was this what you try to prevent here? won't work though.

No. I've had crashes here, and although I never managed to figure out exactly why, a null pointer dereference seems the "logical" explanation (and the easiest thing to check against).

Shouldn't have found its way into this RR, though.

> On Dec. 21, 2014, 10:41 p.m., Milian Wolff wrote:
> > vcs/vcspluginhelper.cpp, line 206
> > <https://git.reviewboard.kde.org/r/120420/diff/4/?file=333845#file333845line206>
> >
> >     again, can this really happen? when? why now, and not before?

same thing, though in this case it's almost impossible to say whether the crashes I saw were related or not to the issue at hand.
When and why, I have no idea. As I said, I haven't been able to debug the crashes post-mortem, and it's not an option to run the IDE with full debug and zero optimisation for something that happens only occasionally and without a well-defined trigger.

> On Dec. 21, 2014, 10:41 p.m., Milian Wolff wrote:
> > vcs/vcspluginhelper.cpp, line 214
> > <https://git.reviewboard.kde.org/r/120420/diff/4/?file=333845#file333845line214>
> >
> >     also looks wrong. the list should never contain zero items. if it does, something else is wrong

yes, I'd hope the null pointer check was omitted with good reason.

This (and the issues above) is one of those cases where I'm banging my head against the impossibility to check a pointer's validity without resorting to expensive things like exceptions or signal trapping. The fact something went wrong somewhere leading to weird arguments elsewhere doesn't mean the code shouldn't try to exit gracefully...

- René J.V.

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/120420/#review72382
-----------------------------------------------------------

On Dec. 19, 2014, 6:52 p.m., René J.V. Bertin wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://git.reviewboard.kde.org/r/120420/
> -----------------------------------------------------------
> 
> (Updated Dec. 19, 2014, 6:52 p.m.)
> 
> 
> Review request for KDE Software on Mac OS X and KDevelop.
> 
> 
> Repository: kdevplatform
> 
> 
> Description
> -------
> 
> This RR replaces https://git.reviewboard.kde.org/r/120150/, itself a follow-up to RR https://reviewboard.kde.org/r/120081/ . In that RR I proposed an (accepted & submitted ) approach to prevent kdevelop from crashing after closing the patch review ("git/show differences") toolview.
> 
> This turned out to be insufficient, the most likely culprit being a nested event loop. In my most recent bug hunting I noticed that `closeReview` is called through Qt's mouse-eventh handling function when the "Finish Review" button (or "Commit", when doing a git->commit) is pressed. The crash I have been experiencing occurs in that function (or at least not far down the call stack from that function).
> I had already noticed that the crash never occurred when I closed all document views first (the patch file and all modified files), e.g. by using the "Close All" menu action.
> 
> So:
> - `closeReview` closes all open documents (minus the patchfile) via `qDeleteAll(m_highlighters)` or by calling `m_highlighters.erase(iterator)` repeatedly. Documents are thus closed by deleting the "content object" that represents them, letting that fact filter back to the UI
> - closing documents via the UI takes the opposite route, and is more in line with a modern GUI framework (at least OS X/Cocoa) where everything happens because of an event in the user interface. You receive an event (message) when the user closes a document, the framework handles most UI-related stuff for you, and if you have to dispose of document-related non-gui data you can do that for example just before the widget actually closes (`windowWillClose`, `menuWillClose`, `applicationWillClose` etc. messages in Cocoa).
> 
> This led to the current patch, which cycles through and closes the `Sublime::View`s associated with the window area, and then lets Qt send and process all events that are pending and/or result from closing those windows. Only then does it proceed the regular course of action (calling `removeHighLighting()` which ought to be no longer necessary).
> 
> The only change for the user is that the pinkish patch review plugin background (with the "back to code" button) is visible briefly before the toolview is closed.
> 
> I understand that this crash (or a closely related one) does not only occur on OS X but also on Linux, so I did not put the new code in a conditional block.
> 
> 
> Diffs
> -----
> 
>   plugins/grepview/grepviewplugin.cpp 7e3d933 
>   plugins/patchreview/patchreview.cpp 18b63db 
>   shell/sessioncontroller.cpp ae4e355 
>   vcs/vcspluginhelper.cpp 26ab57c 
>   vcs/widgets/vcsdiffpatchsources.cpp c3e2dae 
>   vcs/widgets/vcsdiffwidget.cpp 54787b9 
> 
> Diff: https://git.reviewboard.kde.org/r/120420/diff/
> 
> 
> Testing
> -------
> 
> kdevplatform git/kde4-legacy on KDE/MacPorts OS X 10.6.8 with kdelibs 4.14.1
> 
> 
> Thanks,
> 
> René J.V. Bertin
> 
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/kdevelop-devel/attachments/20141221/f31b14ca/attachment.html>