Fixing a crash when a game offers both to create a rematch and to delete
itself because its opponent has been deleted. Deleting and then hitting
rematch would crash because the rematch alert referred to a dead
game. Now onStop() for delegates removes any pending alerts. In
dual-pane mode only since when there's an activity going away the alerts
go automatically.
Rather than have callers of getSummary() try JNIThread for the lock, do
that check inside getSummary(), and move it to GameUtils from DBUtils
since it's using higher-level knowledge now.
Add new parameterless setTitle() method on delegates, and call it when a
fragment is removed so the new right-most pane can restore a title that
makes sense. So far only board and gameconfig delegates implement this
new method.
Current GameLock implementation means you can't get a lock for an open
game, so try getting one from an existing JNIThread instance
first. Which is a hack that's start to appear in lots of places.... Also
fix so just in case we are unable to lock a game we drop the rematch
process rather than crash in an assert later. The test case: rematch a
solo game that's currently open in the right pane.
Fixed race conditions revealed by dual-pane mode where GameLock could be
instantiated and then attempts made to reference its game (jni calls)
before it had been loaded. So now loading happens inside the same
synchronized methods as opening or creating a game.
Rare crashes are happening inside the jni, in game_dispose(), when a
game's double-disposed. Adding a jni call to check if the thread about
to do a game_dispose() will fail then asserting its result in java
allows useful stack traces to come via Crittercism. Or should.
There was a race condition between finish() and the popping of fragments
that happens inside dispatchNewIntent(). If dispatchNewIntent() won then
later finish() would pop the GamesListFragment and we'd crash. Ideally
finish() would pass a fragment to finishFragment() which would then do
nothing if that fragment wasn't on the stack. Later....
Trying and failing to get just the content area of the tree view to
scroll, whether by putting it in its own container or hooking into the
scrollable interface tree view allegedly supports.
better strings, and explain when pref changed that user must restart for
it to take effect. Actually restarting from inside prefs delegate is
hard enough I'm not doing it for what should be advanced users.
I want to use ReentrantLock instead of my implementation but I'm
breaking its rule that the thread that locks a lock must be the one to
unlock it. Add commented-out assertion for some later time when I might
want to fix this. No change for now.
Game created for rematch was coming up unconnected and without
explanation if recipient of invitation hadn't responded. Don't dismiss
the INVITE alert in that case.
This to prevent the game list item from resizing (and so reflowing the
whole list) when it's reloading game info. Better would be to set a
timer and only show the reloading view if it seems likely to take a
minute.
Toward something that should work with android: pass a potentially
unique draw context into new method that creates a new board just for
the draw and makes its scoreboard and tray disappear.
Done in a way that won't work for Android just to try out the gtk
way. Tweaks and reuses the existing draw context and board, which is
precisely what the android version can't easily do.
Add new "snap" blob to db and pixbuf column to games display table, and
add code to store and retrieve from db. What's stored now is a
hard-coded file, so next up is generating an actual snapshot from the
game.
When a gtk3 window's shutting down it appears we can't get a cairo_t*
for it. This change makes it possible to turn that fact into aborting
the whole draw operation.
Whenver we get the draw signal, invalidate the whole board and draw
it. This is resulting in a lot of wasted effort that might be mitigated
if I figured out how to get the invalid region out of the cairo_t*
that's passed in, but this is a test app and my machine's pleny fast. So
go with it for now.
I'm seeing unreproducible crashes trying to double-dispose jni game
instances and think it's a race condition involving JNIThread. Forcing
it to hold the lock in the constructor, access to which is synchronized,
is an improvement and may well fix it.
Board renders, but only when touched. Tray and scoreboard skipped for
now. Lots of changed still to go, and some asserts added where I didn't
want to stop to solve a compile problem.
When a device is a tablet and not a first-time install, put up an offer
to enable dual-pane mode. Change confirm-alerts to include
do-not-show-again box, and use that. Add menu item, hidden when not in
dual-pane mode, to turn it back off. Exit app after posting a
notification and a toast on changing that preference so it'll take
effect.
When app's launched via a move-made intent that will lead to opening
the board, let the GamesList fragment get fully in place before
opening the board. To open two at the same time confuses my fragment
code (OS kills it with a fatal exception.)
When during onResume() of BoardDelegate we notice an undisplayed chat
message, don't add add a Chat fragment because (perhaps due to a bug
in my code) it'll come up blank. Instead use post() to open it via a
Runnable() that'll run after onResume() and the rest of current
fragment setup have completed.
When loading gamelistitems in dual-pane mode there will often be an
open game. Don't hold up the whole process by first waiting 1 second
to get a lock that's unavailable. Instead check if there's a JNIThread
instance available and if so use its lock to get the summary. Required
fixing JNIThread to not crash trying to save when released too early.
Fixes, or at least makes extremely unlikely, race condition where one
thread makes use of an existing JNIThread instance then releases it
before the thread that created it has had a chance to call configure()
to actually load the game.
Problem was that changes to games didn't show up in the thumbnail
until the game was closed. Simply using existing snapshot didn't work
because it changes the board layout in order to "draw" to
ThumbnailCanvas and that change isn't easily reversed. So copy
existing code to open a new JNI object with just-saved game data
and use it to create a thumbnail bitmap.
They're often needed when the fragment isn't frontmost, i.e. when
onPause() has been called. My caching instances fix is feeling a bit
fragile, but I think it's better than nothing. Alternative is probably
to go with DialogFragments, a big change that might not be easy to
make work back to oldest Android.
There's a problem in dual-pane mode where activites outlive
DelegateBase instances that are tied to fragments. AlertDialogs, being
bound to the MainActivity, can sometimes outlive the delegates that
create them, meaning the 'this' referred to from closures bound to
onClick() handlers can come to be invalid (e.g. referencing a removed
fragment). So add a global registry of current DelegateBase instances
by class, and from onClick() handlers fetch and use the current
instance instead of the 'this' that's bound.
Inviting didn't work because it's done by a separate activity whose
onActivityResult() was dropped because DualpaneDelegate was the
recipient. That now handles it by asking MainActivity to sent it to a
contained Delegate. Currently will go only to the top (rightmost) one.
Was opening game in init(), but in dual-pane case there's no
onWindowFocusChanged() call from which to check state, so open it
instead from onResume(), at which point things are already in place to
handle callbacks immediately. That is, post() will work.
use s/getTargetFragment(), but there's still the hack of capturing
state in setResult() and then invoking fragment.onActivityResult()
blindly the next time the backstack is popped.
create queue of Runnables to be run only when Delegate is
visible (onResume() has been called and onPause() hasn't.) When
missing dict finishes downloading, rather than immediately opening its
game, post a Runnable to do it that will be run only after the main
activity is ready to have fragment transactions committing again.
When the game's already opened references must be obtained this
way. This may not be necessary however once game config is opened on
top of games list instead of on top of the game as it should be.
For some reason the 'clean-debug' cycle stopped working because the
file was nuked then not regenerated. Rather than figure out why, just
add a target to generate it every time.
On 4.2 and 4.3 at least, the dimiss comes after the second post
causing that to be dismissed, leaving no alert and the app in some
weird state where not even the back button works. It's been this way
on those older versions since I redid invitations for beta 98.
