QGIS » QGIS Application

QGIS 2.12 causes a crash when closing the 'DB Manager' in the 'Preview' view (Win7).

QGIS 2.12 causes a minidump when closing the 'DB Manager' in the 'Preview' view with vector postgis (Win10) (see Attachment)

QGIS 2.12 causes a minidump when closing the 'DB Manager' in the 'Preview' view with vector postgis (Win10) (see Attachment)

http://1drv.ms/1XUuSyS

salvatore fiandaca wrote:

QGIS 2.12 causes a minidump when closing the 'DB Manager' in the 'Preview' view with vector postgis (Win10) (see Attachment)

http://1drv.ms/1XUuSyS

I just tried on the latest master (and 2.12 too) and does not happen, could you try it too (and remove/disable all the plugins too)?

I started with --no plugins; I can not find DBManager

salvatore fiandaca wrote:

I started with --no plugins; I can not find DBManager

sorry, I meant 3rd party plugins.

salvatore fiandaca wrote:

remove all 3rd party plugins, the mini-dump remains.!!!!

can you replicate on other machines?

I installed QGIS 2.12 in another machine (6 GB RAM, 64-bit, Win 10, like mine), the mini-dump remains.

It occurred to me that preventing the mapcanvas from rendering might help.
have found that in db_manager.py, def closeEvent(self,e) using:

    self.unregisterAllActions()
    self.preview.freeze(True)       #<-- Added line: preview (LayerPreview) is derived from QgsMapCanvas
    self.preview.loadPreview(None)

Further testing with rev. 93bd825 the crash occurs on 32-bit Windows builds, but doesn't happen as often (for a while, not on 10 or more tries) on 64-bit; adding the freeze() line stops the crash on both.

Might there a problem with passing references across threads somewhere in the map rendering similar to the Browser crashing (see #13738)?

Using setRenderFlag(False) instead of freeze() does not prevent crash -- why?

Further investigation of QgsMapCanvas -- freeze() and setRenderFlag() -- Both do the same thing, which is to allow/disallow rendering the map. However, the setRenderFlag() function does something in addition. The mRenderFlag is tested and if false calls stopRendering() [and refresh() if true -- not relevant to this problem]

So using freeze(True) is a better choice.

Steven Mizuno wrote:

It occurred to me that preventing the mapcanvas from rendering might help.
have found that in db_manager.py, def closeEvent(self,e) using:

does this fixes this issue that I can't confirm? if yes then if possible make a patch and pull request. Thanks!

Giovanni Manghi wrote:

salvatore fiandaca wrote:

I installed QGIS 2.12 in another machine (6 GB RAM, 64-bit, Win 10, like mine), the mini-dump remains.

do you still have issue? I cannot really replicate this one in my computer and any other one. I have just gave a postgis training to a classroom full of people (all Windows) and no one had this issue.

yes, I always have the problem. Patience.

I cannot reproduce this on maste (a0e6d49) with ubuntu-14.04 -- any chance you can re-test this Salvatore ?

RuntimeError: wrapped C/C++ object of type QgsVectorLayer has been deleted 
Traceback (most recent call last):
  File "/usr/src/qgis/build/master/output/python/plugins/db_manager/db_manager.py", line 62, in closeEvent
    self.preview.loadPreview(None)
  File "/usr/src/qgis/build/master/output/python/plugins/db_manager/layer_preview.py", line 57, in loadPreview
    self._clear()
  File "/usr/src/qgis/build/master/output/python/plugins/db_manager/layer_preview.py", line 85, in _clear
    self._loadTablePreview(None)
  File "/usr/src/qgis/build/master/output/python/plugins/db_manager/layer_preview.py", line 118, in _loadTablePreview
    QgsMapLayerRegistry.instance().removeMapLayers([self.currentLayer.id()])
RuntimeError: wrapped C/C++ object of type QgsVectorLayer has been deleted

Sandro,
il minidump mi perseguita, sia se utilizzo db postgis che spatialite.
poi ho riscontrato, (sempre su dbmanager) altri problemi legati a file csv; cioè se carico in qgis un file csv con CTRL+SHIFT+V e poi cerco di importarlo in un db spatialite tramite dbmanager mi da errore ' errore 5 errore nella creazione dei campi'.

ciao

The call to loadPreview on close() was added by Giuseppe to fix #12938
Giuseppe, any idea about my backtrace ?

Salvatore please use different tickets for different bugs. The fact they all result in a "minidump" do not make them be the same bug.

I cannot always reproduce it, btw

Actually, I cannot reproduce it anymore :/

il minidump si verifica solo ed esclusivamente se facciamo 'anteprima' del layer (Postgis e spatialite);
infatti evito di fare l'anteprima per lavorare. (QGIS 2.12.3)

the mini-dump occurs if and only if I 'preview' layer (Postgis and spatialite);
In fact, it avoids doing the preview to work. (QGIS 2.12.3)

No crash here (Ubuntu 14.04 64bit), but in log panel I can see the following messages:

2016-01-18T19:58:26 1 QObject::connect: Cannot connect (null)::repaintRequested() to LayerPreview::refresh()
2016-01-18T19:58:26 1 QObject::connect: Cannot connect (null)::layerCrsChanged() to LayerPreview::layerCrsChange()
2016-01-18T19:58:28 1 Loading a style file that was saved with an older version of qgis (saved in 2.11.0-Master, loaded in 2.13.0-Master). Problems may occur.
2016-01-18T19:58:28 1 QObject::connect: Cannot connect (null)::repaintRequested() to LayerPreview::refresh()
2016-01-18T19:58:28 1 QObject::connect: Cannot connect (null)::layerCrsChanged() to LayerPreview::layerCrsChange()

that is similar to the issue reported by Sandro (strk): the layer was deleted.

In my case it seems due to a canvas refresh (repaintRequested), so I guess it's a race-condition due to multithread (but I've never looked at the current implementation, so I cannot say anything more).

Surely the issue can occurs if the layer is destroyed using delete (and not deleteLater() which ensure signal disconnection).

Since objects in python are destroyed when the last ref variable goes out of scope (not by a deferred deletion event), something wrong can happen.

Is there any way to make the python wrapper call deleteLater rather than delete, on dropping last reference ?

In the meantime, freezing preview canvas and/or change the selected tab (hiding preview) on close could mitigate the issue since the canvas is not repainted, but I haven't tested it.

I still cannot reproduce with b2a1273.
Could it have been done to the recent threading fixes ?
Which version did you try with, Giuseppe ?
Also, does having or not to enter a password make a difference ?

I have tried rev. 81448f8 (2016-01-19) and DB Manager closing still crashes QGIS if a layer is previewed. This is using 32-bit Windows build on Win8 or Win7 64-bit systems.

I do see the following messages when the Preview tab is clicked for a layer. These occur when the layer is shown, not when DB Manager is closed, so I don't know if they are relevant.

2016-01-20T19:49:13    1    QObject::connect: Cannot connect (null)::repaintRequested() to LayerPreview::refresh()
2016-01-20T19:49:13    1    QObject::connect: Cannot connect (null)::layerCrsChanged() to LayerPreview::layerCrsChange()

As I previously reported this occurs on SpatiaLite layers as well as PostGIS, so I believe that a password is NOT a factor (see comment 31 and the subject).

From my notes about the problem I found that I captured a call stack trace using WinDbg -- this was from an earlier rev., however. The top line is the last call logged before the crash.

QtGui4!QGraphicsItem::prepareGeometryChange+0x9
qgis_gui!QgsMapCanvasMap::setContent(class QImage * image = 0x004ec4c0, class QgsRectangle * rect = 0x004ec38c)+0x13d
qgis_gui!QgsMapCanvas::rendererJobFinished(void)+0xbd5
qgis_gui!QgsMapCanvas::qt_static_metacall(class QObject * _o = 0x042ac468, QMetaObject::Call _c = InvokeMetaMethod (0n0), int _id = 0n40, void ** _a = 0x004ec5a4)+0x3f0
QtCore4!QMetaObject::activate+0x3c6
qgis_core!QgsMapRendererJob::finished(void)+0x1a
qgis_core!QgsMapRendererSequentialJob::internalFinished(void)+0x15b
qgis_core!QgsMapRendererSequentialJob::qt_static_metacall(class QObject * _o = 0x043ff0f8, QMetaObject::Call _c = InvokeMetaMethod (0n0), int _id = 0n0, void ** _a = 0x004ec684)+0x2d
QtCore4!QMetaObject::activate+0x3c6
qgis_core!QgsMapRendererJob::finished(void)+0x1a
qgis_core!QgsMapRendererCustomPainterJob::futureFinished(void)+0xaf
qgis_core!QgsMapRendererCustomPainterJob::cancel(void)+0x22a
qgis_core!QgsMapRendererSequentialJob::cancel(void)+0x9f
qgis_gui!QgsMapCanvas::~QgsMapCanvas(void)+0x16cqgis_gui!QgsMapCanvas::~QgsMapCanvas(void)+0x16c
_gui!sipQgsMapCanvas::~sipQgsMapCanvas(void)+0x6b
_gui!sipQgsMapCanvas::`scalar deleting destructor'(void)+0xf
QtCore4!QObjectPrivate::deleteChildren+0x72
QtGui4!QWidget::~QWidget+0x5c3
QtGui4!QWidgetPrivate::invalidateBuffer_resizeHelper+0x8af
QtGui4!QRegion::~QRegion+0xb
QtGui4!QStackedWidget::~QStackedWidget+0x7a

Does 09ce8004fd2ca9e911b84dd4de59bafb47037b44 (from Salvatore Larosa) fix the crash ?
Giuseppe: wouldn't it be better to avoid calling a "refresh" function on "close" ? It does sound like an hack...

Salvatore Larosa has solved the mini-dump. Hooray!