cascading failure causing segfault in GtkReactive tests?

[vtjnash]

https://s3.amazonaws.com/julialang-reports/nanosoldier/pkgeval/by_date/2021-02/11/GtkReactive.1.7.0-DEV-50742c7e42.log

(julia:170): GLib-GObject-CRITICAL **: 06:21:42.833: g_object_ref_sink: assertion 'G_IS_OBJECT (object)' failed

(julia:170): GLib-GObject-CRITICAL **: 06:21:42.843: g_object_setv: assertion 'G_IS_OBJECT (object)' failed

(julia:170): GLib-GObject-CRITICAL **: 06:21:42.872: g_closure_new_object: assertion 'G_IS_OBJECT (object)' failed

signal (11): Segmentation fault
in expression starting at /home/pkgeval/.julia/packages/GtkReactive/gQSXq/test/runtests.jl:22
unsafe_store! at ./pointer.jl:118 [inlined]
_signal_connect at /home/pkgeval/.julia/packages/Gtk/TgEIW/src/GLib/signals.jl:40
signal_connect at /home/pkgeval/.julia/packages/Gtk/TgEIW/src/GLib/signals.jl:27 [inlined]
signal_connect at /home/pkgeval/.julia/packages/Gtk/TgEIW/src/GLib/signals.jl:27

[tknopp]

Is this a) reproducible and b) also happens in 1.6? I also have seen segfaults on 1.6-beta on an internal software of mine. A minimal reproducible example would be good to track this down.

[giordano]

Those are the tests run by PkgEval. Note that tests for nightly have been been failing forever since we switched to Github Actions: https://github.com/JuliaGraphics/Gtk.jl/actions?query=workflow%3ACI

[tknopp]

Hm, yes. But nightly != 1.6 right?

[giordano]

Yes, but I was referring to point 1, tests have been consistently failing for months. Maybe looking at them might help. Julia v1.6 isn't tested at the moment in CI (but you just need to add one line to the matrix)

[tknopp]

Hm, "Error: Could not find a Julia version that matches 1.6". Do you know what to enter there? 1.6-?

[tknopp]

tests passing locally on Julia 1.6rc1

[giordano]

At the moment it's ^1.6.0-0, it'll be 1.6.0, 1.6, or 1 when it'll be released

[tknopp]

radio: Error During Test at /home/runner/work/Gtk.jl/Gtk.jl/test/gui.jl:340
170
  Test threw exception
171
  Expression: get_gtk_property(get_gtk_property(r, :active), :label, AbstractString) == choices[2]
172
  no active elements in GtkRadioGroup

seems unrelated to the segfault though.

[tknopp]

ok, its the same test failure as on nightly.

[tknopp]

I do not get why this is failing on Julia 1.6 (CI) but not not on 1.3 (CI) and also not locally on 1.6.

[tknopp]

Locally I get:

julia> get_gtk_property(get_gtk_property(r,:active),:label,AbstractString)
"choice two"

[tknopp]

@giordano: Can you somehow identify if different Gtk binaries are used for different Julia versions?

[giordano]

also not locally on 1.6.

That's the fun part of running CI, isn't it? 😃

Note that you can SSH into the github actions runners for debugging with the step

    - name: Setup tmate session
      uses: mxschmitt/action-tmate@v3
      with:
        limit-access-to-actor: true

right before the run tests step (you can remove the last two lines if you don't want to use your ssh key to login). See https://github.com/marketplace/actions/debugging-with-tmate

[timholy]

TIL, thanks @giordano!

[tknopp]

Ok, I have some findings and can even reproduce the issue locally. Here is my test program:

# file is called gui2.jl here locally
using Test, Gtk
using Gtk.ShortNames, Gtk.GConstants, Gtk.Graphics

numFailures = 0
for l=1:1000
  choices = ["choice one", "choice two", "choice three", RadioButton("choice four"), Label("choice five")]
  r = RadioButtonGroup(choices,2)
  active = [get_gtk_property(b,:active,Bool) for b in r]
  if sum(active) != 1
    @show active
    global numFailures += 1
  end
end

@show numFailures

So the issue is with the calls where the vector of active elements is obtained. Sometimes (not always) I get the following run:

julia> include("gui2.jl")

(julia:6981): GLib-GObject-CRITICAL **: 20:43:05.590: g_object_ref_sink: assertion 'G_IS_OBJECT (object)' failed

(julia:6981): GLib-GObject-CRITICAL **: 20:43:05.596: g_object_get_property: assertion 'G_IS_OBJECT (object)' failed

(julia:6981): GLib-GObject-CRITICAL **: 20:43:05.596: g_object_ref_sink: assertion 'G_IS_OBJECT (object)' failed

(julia:6981): GLib-GObject-CRITICAL **: 20:43:05.596: g_object_get_property: assertion 'G_IS_OBJECT (object)' failed

(julia:6981): GLib-GObject-CRITICAL **: 20:43:05.596: g_object_ref_sink: assertion 'G_IS_OBJECT (object)' failed

(julia:6981): GLib-GObject-CRITICAL **: 20:43:05.596: g_object_get_property: assertion 'G_IS_OBJECT (object)' failed
active = Bool[0, 0, 0, 0, 0]
numFailures = 1
1

So it seems that something went wrong Gtk internally and then simply false is returned even for the active element.

[tknopp]

I do not get this error with Julia 1.5.

[tknopp]

So these sort of spontaneous errors are not the one I am capable of tracking down since this certainly touches the core of the type conversion in done in the glib module. The interesting question is what has changed from Julia 1.5 to 1.6 that can lead to these non-deterministic sort of errors. Ping @vtjnash who is much better prepared to solve this and ping @KristofferC to make you aware of this regression.

[jonathanBieler]

I suspect it's an issue with custom widgets since RadioButtonGroup is one (see also #557). Maybe there's some issues here ?

Gtk.jl/src/GLib/gtype.jl

Line 489 in b61cb23

function gobject_move_ref(new::GObject, old::GObject)

[timholy]

@jonathanBieler, the RadioButtonGroup code doesn't use gobject_move_ref:

Gtk.jl/src/buttons.jl

Line 64 in d743a8b

GtkRadioButtonGroup(layout::GtkContainer) = new(layout)

If I change the rhs to gobject_move_ref(new(layout), layout), I get this:

(julia:85406): GLib-GObject-CRITICAL **: 07:54:00.060: g_object_set_qdata_full: assertion 'G_IS_OBJECT (object)' failed

(julia:85406): GLib-GObject-CRITICAL **: 07:54:00.060: g_object_ref_sink: assertion 'G_IS_OBJECT (object)' failed

(julia:85406): Gtk-CRITICAL **: 07:54:00.075: gtk_container_add: assertion 'GTK_IS_CONTAINER (container)' failed

(julia:85406): Gtk-CRITICAL **: 07:54:00.075: gtk_container_add: assertion 'GTK_IS_CONTAINER (container)' failed

(julia:85406): Gtk-CRITICAL **: 07:54:00.076: gtk_container_add: assertion 'GTK_IS_CONTAINER (container)' failed

(julia:85406): Gtk-CRITICAL **: 07:54:00.103: gtk_container_add: assertion 'GTK_IS_CONTAINER (container)' failed

(julia:85406): Gtk-CRITICAL **: 07:54:00.118: gtk_container_add: assertion 'GTK_IS_CONTAINER (container)' failed

all immediately upon calling RadioButtonGroup(choices,2).

EDIT: the assertion 'G_IS_OBJECT (object)' failed are triggered at

Gtk.jl/src/GLib/gtype.jl

Line 494 in b61cb23

gc_ref(new)

and the assertion 'GTK_IS_CONTAINER (container)' failed by each push!:

Gtk.jl/src/buttons.jl

Line 74 in d743a8b

push!(grp, e, i == active)

[timholy]

OK, this diff gets us closer to the design suggested in https://juliagraphics.github.io/Gtk.jl/latest/manual/customWidgets/:

diff --git a/src/buttons.jl b/src/buttons.jl
index 4574ee9..e19b533 100644
--- a/src/buttons.jl
+++ b/src/buttons.jl
@@ -59,9 +59,9 @@ mutable struct GtkRadioButtonGroup <: GtkContainer # NOT a native @gtktype
     # the behavior is specified as undefined if the first
     # element is moved to a new group
     # do not rely on the current behavior, since it may change
-    handle::GtkContainer
+    handle::Ptr{GObject}
     anchor::GtkRadioButton
-    GtkRadioButtonGroup(layout::GtkContainer) = new(layout)
+    GtkRadioButtonGroup(layout::GtkContainer) = gobject_move_ref(new(unsafe_convert(Ptr{GObject}, layout)), layout)
 end
 const GtkRadioButtonGroupLeaf = GtkRadioButtonGroup
 macro GtkRadioButtonGroup(args...)
@@ -88,7 +88,7 @@ function push!(grp::GtkRadioButtonGroup, e::GtkRadioButton)
     else
         grp.anchor = e
     end
-    push!(grp.handle, e)
+    container_add!(grp.handle, e)
     grp
 end
 function push!(grp::GtkRadioButtonGroup, label, active::Union{Bool, Nothing} = nothing)
@@ -100,7 +100,7 @@ function push!(grp::GtkRadioButtonGroup, label, active::Union{Bool, Nothing} = n
     if isa(active, Bool)
         gtk_toggle_button_set_active(e, active::Bool)
     end
-    push!(grp.handle, e)
+    container_add!(grp.handle, e)
     grp
 end
 function start_(grp::GtkRadioButtonGroup)
diff --git a/src/container.jl b/src/container.jl
index a07ee41..b6609ee 100644
--- a/src/container.jl
+++ b/src/container.jl
@@ -1,9 +1,11 @@
+container_add!(w, child)    = ccall((:gtk_container_add, libgtk),    Nothing, (Ptr{GObject}, Ptr{GObject},), w, child)
+container_remove!(w, child) = ccall((:gtk_container_remove, libgtk), Nothing, (Ptr{GObject}, Ptr{GObject},), w, child)
 function push!(w::GtkContainer, child)
-    ccall((:gtk_container_add, libgtk), Nothing, (Ptr{GObject}, Ptr{GObject},), w, child)
+    container_add!(w, child)
     w
 end
 function delete!(w::GtkContainer, child::GtkWidget)
-    ccall((:gtk_container_remove, libgtk), Nothing, (Ptr{GObject}, Ptr{GObject},), w, child)
+    container_remove!(w, child)
     w
 end
 function empty!(w::GtkContainer)

Sometimes it succeeds, but I still get intermittent failures like

(julia:92982): GLib-GObject-CRITICAL **: 08:57:46.379: g_object_ref_sink: assertion 'G_IS_OBJECT (object)' failed

(julia:92982): GLib-GObject-CRITICAL **: 08:57:46.384: g_object_get_property: assertion 'G_IS_OBJECT (object)' failed

that seem to be reduced by adding various sleeps. What's the right way to do this?

[vtjnash]

I think you might need a GC.@preserve w in places like

Gtk.jl/src/buttons.jl

Line 115 in d743a8b

iterate(w::GtkRadioButtonGroup, s=start_(w)) = iterate(s, s)

[timholy]

Ho ho! I hadn't noticed that, but that's a good catch.

I added it everywhere start_ gets used in buttons.jl (there are other uses for other types but I didn't try it there). I still get intermittent errors like the above.

[timholy]

I think you're seriously on to something, though: if I wrap

Gtk.jl/test/gui.jl

Lines 314 to 342 in d743a8b

	choices = ["choice one", "choice two", "choice three", RadioButton("choice four"), Label("choice five")]
	w = Window("Radio")
	f = Gtk.GtkBox(:v); push!(w,f)
	r = Vector{RadioButton}(undef, 3)
	r[1] = RadioButton(choices[1]); push!(f,r[1])
	r[2] = RadioButton(r[1],choices[2]); push!(f,r[2])
	r[3] = RadioButton(r[2],choices[3],active=true); push!(f,r[3])
	@test [get_gtk_property(b,:active,Bool) for b in r] == [false, false, true]
	set_gtk_property!(r[1],:active,true)
	@test [get_gtk_property(b,:active,Bool) for b in r] == [true, false, false]
	showall(w)
	destroy(w)
	r = RadioButtonGroup(choices,2)
	@test length(r) == 5
	@test sum([get_gtk_property(b,:active,Bool) for b in r]) == 1
	itms = Vector{Any}(undef,length(r))
	for (i,e) in enumerate(r)
	itms[i] = try
	get_gtk_property(e,:label,AbstractString)
	catch
	e[1]
	end
	end
	@test setdiff(choices, itms) == [choices[4],]
	@test setdiff(itms, choices) == ["choice four",]
	@test get_gtk_property(get_gtk_property(r,:active),:label,AbstractString) == choices[2]
	w = Window(r,"RadioGroup")|>showall
	destroy(w)

in GC.enable(false); ... ; GC.enable(true) it seems to pass every time.

[timholy]

There really does seem to be something different between 1.5 and 1.6. If you put

GC.gc(true); GC.gc(true); GC.gc(true)

right after the r = RadioButtonGroup(choices,2), then on 1.6 I get failures on alternating runs. (weird) It never seems to fail on 1.5.

As disclosure, I've also commented out most steps that run prior to the "radio" tests, and I've also removed the @testset.

diff --git a/test/gui.jl b/test/gui.jl
index 024af93..e5a9b81 100755
--- a/test/gui.jl
+++ b/test/gui.jl
@@ -2,7 +2,7 @@
 
 using Gtk.ShortNames, Gtk.GConstants, Gtk.Graphics
 import Gtk.deleteat!, Gtk.libgtk_version, Gtk.GtkToolbarStyle, Gtk.GtkFileChooserAction, Gtk.GtkResponseType
-
+#=
 ## for FileFilter
 # This is just for testing, and be careful of garbage collection while using this
 struct GtkFileFilterInfo
@@ -309,8 +309,8 @@ set_gtk_property!(check,:label,"new label")
 showall(w)
 destroy(w)
 end
-
-@testset "radio" begin
+=#
+# @testset "radio" begin
 choices = ["choice one", "choice two", "choice three", RadioButton("choice four"), Label("choice five")]
 w = Window("Radio")
 f = Gtk.GtkBox(:v); push!(w,f)
@@ -325,7 +325,9 @@ showall(w)
 destroy(w)
 
 r = RadioButtonGroup(choices,2)
+GC.gc(true); GC.gc(true); GC.gc(true)
 @test length(r) == 5
+GC.enable(false)
 @test sum([get_gtk_property(b,:active,Bool) for b in r]) == 1
 itms = Vector{Any}(undef,length(r))
 for (i,e) in enumerate(r)
@@ -335,12 +337,17 @@ for (i,e) in enumerate(r)
             e[1]
         end
 end
+GC.enable(true)
 @test setdiff(choices, itms) == [choices[4],]
 @test setdiff(itms, choices) == ["choice four",]
+GC.enable(false)
 @test get_gtk_property(get_gtk_property(r,:active),:label,AbstractString) == choices[2]
+GC.enable(true)
 w = Window(r,"RadioGroup")|>showall
 destroy(w)
-end
+# end
+# GC.enable(true)
+error("stop")
 
 @testset "ToggleButton" begin
 tb = ToggleButton("Off")

[timholy]

Another observation: if I combine the above with disabling GC while the RadioButtonGroup constructor runs, then I get the assertion 'G_IS_OBJECT (object)' failed on every run. So it seems that GC needs to run during construction?

[timholy]

I am now onto other deadlines, but to leave this in a good state, here's a MWE:

using Revise   # curiously, this seems to be important to making this fully reproducible, though CI fails without
using Gtk, Gtk.ShortNames, Test

choices = ["choice one", "choice two", "choice three", RadioButton("choice four"), Label("choice five")]
r = RadioButtonGroup(choices,2)
GC.gc(true); GC.gc(true); GC.gc(true)
@test length(r) == 5
@test sum([get_gtk_property(b,:active,Bool) for b in r]) == 1

This reliably passes on 1.5 and reliably fails on 1.6. If you comment out the GC.gc(true); GC.gc(true); GC.gc(true) line then it seems to pass most or all the time on 1.6. The outcome is repeatable even in the same session, you don't need a restart.

Both are running Gtk 1.1.6.

[timholy]

I've opened an issue (see link above) that identifies JuliaLang/julia#38180 as the probable source of the change. What's really odd is that Gtk doesn't use WeakKeyDict (lots of WeakRefs though), but there were also changes to gc.c.

[vtjnash]

That changed WeakRef so they must be re-set by a finalizer if it wants to keep them active. E.g.

w = WeakRef(x)
finalizer(x) do x; keepalive(x) && w.value = x; end

[timholy]

Sorry, I don't understand. I don't find keepalive in Base, GC, Gtk, or Gtk.GLib. Are there any docs on this?

[vtjnash]

I suppose I should have just said the_rest_of_the_finalizer(x)

[timholy]

OK. It would be ideal if someone who understands memory management in Gtk (@jonathanBieler ?) fixes this here, and @vtjnash adds the requisite NEWS entry (JuliaLang/julia#39811 (comment)). If @jonathanBieler can't fix it I can tackle it, but I'd like to wait for @vtjnash's NEWS entry first since the enhanced clarity will save me time (which is in short supply).

[jonathanBieler]

@timholy Sadly I don't understand much about memory management in Gtk.jl or Julia for that matter.

[vtjnash]

I don't see an obvious reason it should behave different. Someone might want to run this under rr, so you watch where the g_object_unref calls happen and see where it should be fixed.

[timholy]

Sorry @jonathanBieler, I had you mixed up with with @bfredl who added much of the glib stuff but who hasn't been around a lot lately.

[timholy]

Unfortunately, every time I've tried to use rr in the last two years I get segfaults that don't happen without it:

$ rr record ./julia-debug --args /tmp/gtk.jl
rr: Saving execution to trace directory `/home/tim/.local/share/rr/julia-debug-0'.
Segmentation fault

[tknopp]

( I know this does not change anything, but a big "thank you" for investigating Tim )