Issue with sigatom

[tknopp]

Dear @vtjnash,

There is still a major limitation of the Julia Gtk bindings compared to other language bindings. Its that one can only do one thing at a time. I get

FATAL ERROR: Gtk state corrupted by error thrown in a callback:
ERROR: AssertionError: g_stack === nothing && !prev
Stacktrace:
 [1] g_sigatom(::Any) at /Users/knopp/.julia/packages/Gtk/ADDrn/src/GLib/signals.jl:165
 [2] macro expansion at /Users/knopp/.julia/packages/Gtk/ADDrn/src/GLib/signals.jl:191 [inlined]
 [3] update2 at /Users/knopp/.julia/dev/Gtk/example/task_issue.jl:22 [inlined]
 [4] (::getfield(Base, Symbol("##436#437")){typeof(update2),Timer})() at ./event.jl:475

if I execute the following test program:

using Gtk.ShortNames
using Dates

b = Box(:v)
win = Window(b,"test")
ent = Entry()
push!(b,ent)

button = Button("ok")
push!(b,button)
Gtk.showall(win)

function update_(::Timer)
  Gtk.@sigatom begin
    set_gtk_property!(ent, :text, "$(Dates.now())")
  end
end
timer = Timer(update_, 0.0, interval=0.1)

signal_connect(button, "clicked") do widget
  Gtk.@sigatom begin
    set_gtk_property!(button, :label, "$(rand(1:10))")
    sleep(1.0)
  end
end

You have to click on the button to see the crash.

The example is of course constructed but appears in a real application of mine. The sleep is actually a computation that takes some time.

It would be absolutely awesome if we could find a solution to this problem. Do you think JuliaLang/julia#13099 would solve this, if I just remove sigatom?

[jonathanBieler]

I think that's the correct way to do that:

using GtkExtensions
using Gtk.ShortNames
using Dates

b = Box(:v)
win = Window(b,"test")
ent = Entry()
push!(b,ent)

button = Button("ok")
push!(b,button)
Gtk.showall(win)

function update_cb!(user_data)
    ent = unsafe_pointer_to_objref(user_data)
    set_gtk_property!(ent, :text, "$(Dates.now())")
    Cint(true)#keep running
end

g_timeout_add(100,update_cb!,ent) #do every 100ms

signal_connect(button, "clicked") do widget
    set_gtk_property!(button, :label, "$(rand(1:10))")
end

using this https://github.com/jonathanBieler/GtkExtensions.jl/blob/master/src/GtkExtensions.jl#L337

[tknopp]

If I remove "@sigatom" from something that changes the UI (set*) I get segfaults.

[jonathanBieler]

I never use @sigatom and I never had that kind of issues (on macOS at least). That said this one works fine too:

using GtkExtensions
using Gtk.ShortNames
using Dates

b = Box(:v)
win = Window(b,"test")
ent = Entry()
push!(b,ent)

button = Button("ok")
push!(b,button)
Gtk.showall(win)

function update_cb!(user_data)
    ent = unsafe_pointer_to_objref(user_data)
    Gtk.@sigatom begin
        set_gtk_property!(ent, :text, "$(Dates.now())")
    end
    Cint(true)#keep running
end

g_timeout_add(100,update_cb!,ent) #do every 100ms

signal_connect(button, "clicked") do widget
    Gtk.@sigatom begin
        set_gtk_property!(button, :label, "$(rand(1:10))")
    end
end

[tknopp]

Can you add a sleep to your original example?

[jonathanBieler]

It works fine with a sleep, it blocks the application for 0.1 second though.

[tknopp]

I cannot test your code in Julia 1.0 since cfunction is no more.

[tknopp]

Regarding sleep just try my original example and remove the sigatom, it will segfault. In reality I of course don't have a sleep in my code but it is very hard to prevent a yield if one has a large code base.

[jonathanBieler]

Yes, I think the underlying problem is that it's not safe to interact with Gtk components outside of Gtk's main loop. As I understand sigatom is a workaround, but if you use g_timeout_add and such you'll won't have this problem in the first place since in that case everything is ran from the main loop.

[tknopp]

Could you add that function to Gtk.jl? It is in GLib so it fits.

[jonathanBieler]

I will yes, I have a bunch of stuff in GtkExtensions that should be in Gtk but I haven't had time to move them yet.

Maybe having a look at what other languages do with this would be useful (it seems like Python also has this idle_add).

[tknopp]

yes looking into how other languages do this is a good thing. While my original is a valid use case, what I also would like is to have some heavy computation in the background while the UI is still responsive. For now I have solve this by braking my computation into chunks and using a task for it that always waits between the chunks.

[tknopp]

On julia 1.1 the following code does not segfault anymore:

using Gtk.ShortNames
using Dates

b = Box(:v)
win = Window(b,"test")
ent = Entry()
push!(b,ent)

button = Button("ok")
push!(b,button)
Gtk.showall(win)

function update_(::Timer)
  set_gtk_property!(ent, :text, "$(Dates.now())")
end
timer = Timer(update_, 0.0, interval=0.1)

signal_connect(button, "clicked") do widget
  set_gtk_property!(button, :label, "$(rand(1:10))")
  sleep(1.0)
end

[tknopp]

@vtjnash: I introduced a macro @idle_add that can be used to run code on the UI thread (or simply postpone it when being in a callback). This seems to be the correct solution for Gtk Python and C applications:

https://developer.gnome.org/glib/stable/glib-The-Main-Event-Loop.html#g-timeout-add

After removing @sigatom from my application and replacing it with @idle_add I am wondering if we should simply remove @sigatom or at least not expose it as public API and better point to @idle_add instead. What do you think? @timholy, @jonathanBieler: Whats your take on this? Are you using @sigatom in user code?

[jonathanBieler]

I'm not using @sigatom no, it's also not used in Immerse.jl or GtkUtilities.jl (at least a quick search indicates), so I have no issues with removing it.

[timholy]

This is now coming up in the GtkObservables tests, specifically g_stack is sometimes a Task (and not the current_task()) at

Gtk.jl/src/GLib/signals.jl

Line 174 in 69c1080

@assert g_stack === nothing && !prev

Any ideas of how best to handle this? My knowledge of this part of the code is pretty minimal.

[timholy]

There's a workaround in JuliaGizmos/GtkObservables.jl#30

[tknopp]

We should probably just remove it. The only important location I just found was this:

Gtk.jl/src/events.jl

Line 1 in 543a4f1

gtk_main() = GLib.g_sigatom() do

[timholy]

That seems like a big one, though!

[tknopp]

Yes, indeed. It's the one that wraps everything.