[PATCH 1/2] winex11.drv: Keep the window's own restored state when unminimized, if it was fullscreen before.

[Gabriel Ivăncescu]

Hi Zhiyi,

Thanks for the review.

On 21/10/2020 13:10, Zhiyi Zhang wrote:
> Hi Gabriel,
> 
> Now that the root cause is in winex11.drv, you should add tests in user32.
> For example, minimizing a fullscreen window and then restore it and check its size.
> The d3d9 test doesn't seem to be necessary.
> 

It would still have to be interactive, though, because it has to happen 
within an X11 event (such as clicking on the taskbar). Otherwise the bug 
doesn't manifest. But yeah, I'll try to move it.

> On 10/15/20 10:39 PM, Gabriel Ivăncescu wrote:
>> Fixes a regression introduced by commit
>> 82c6ec3a32f44e8b3e0cc88b7f10e0c0d7fa1b89, which caused the WM_ACTIVATEAPP
>> to be sent while the window is minimized, if it has been clicked on in the
>> taskbar to be restored.
>>
>> According to the Extended Window Manager Hints spec, WMs remove the
>> NET_WM_STATE_FULLSCREEN state when minimized and then, when restored, use
>> the previous size of the window (not the fullscreen) to restore to. This
>> caused the WM_SYSCOMMAND SC_RESTORE message's ShowWindow to revert the window
>> back to its original (non-fullscreen) size, instead of being restored to
>> fullscreen, breaking some apps like Heroes of Might and Magic V.
>>
>> We have to override the X server's window state here to match Windows
>> behavior and restore it back to fullscreen.
>>
>> Signed-off-by: Gabriel Ivăncescu <gabrielopcode at gmail.com>
>> ---
>>
>> Supersedes patches 193752-193754.
>>
>> update_net_wm_states is when Wine does a change and needs to inform the X
>> server about it (i.e. Wine->X11 state change). So, that's good as it is.
>>
>> read_net_wm_states is when X11 changes some state from outside Wine and
>> sends it a notification that it did; then Wine has to update its own wm
>> states to match the X server (so it's properly integrated).
>>
>> However, in this case, since the WM clears the fullscreen state on
>> minimization, we must not do that, because Windows doesn't—and more,
>> when the window is restored from minimized, we must ignore the X11 state
>> changes and force it back to what Wine had (usually fullscreen unless it
>> gets hooked by the app, then we keep what the app changed). This is to
>> match Windows behavior, which differs from X11, and apps rely on.
>>
>>   dlls/winex11.drv/event.c  | 13 +++++++++++++
>>   dlls/winex11.drv/window.c |  9 ++++++++-
>>   dlls/winex11.drv/x11drv.h |  1 +
>>   3 files changed, 22 insertions(+), 1 deletion(-)
>>
>> diff --git a/dlls/winex11.drv/event.c b/dlls/winex11.drv/event.c
>> index 07f7a1a..350d533 100644
>> --- a/dlls/winex11.drv/event.c
>> +++ b/dlls/winex11.drv/event.c
>> @@ -1296,6 +1296,8 @@ static void handle_wm_state_notify( HWND hwnd, XPropertyEvent *event, BOOL updat
>>   
>>       if (data->iconic && data->wm_state == NormalState)  /* restore window */
>>       {
>> +        DWORD old_state = data->net_wm_state;
>> +
>>           data->iconic = FALSE;
>>           read_net_wm_states( event->display, data );
>>           if ((style & WS_CAPTION) == WS_CAPTION && (data->net_wm_state & (1 << NET_WM_STATE_MAXIMIZED)))
>> @@ -1313,11 +1315,22 @@ static void handle_wm_state_notify( HWND hwnd, XPropertyEvent *event, BOOL updat
>>           {
>>               if (style & (WS_MINIMIZE | WS_MAXIMIZE))
>>               {
>> +                /* if the window was fullscreen before minimized, we have to keep its
>> +                   restored state and restore it to its fullscreen state, because WMs
>> +                   remove the fullscreen state when minimized, so don't use their state. */
>> +                if (old_state & (1 << NET_WM_STATE_FULLSCREEN))
>> +                    data->keep_restored_state = TRUE;
> Shouldn't you restrict this path to restore from minimized state only and not from maximized state?
> 
> Do you think 'keep_fullscreen_state' is a more appropriate name?
>   

Might be, I tried to keep the name on the shorter side, and also it's 
possible that the app may hook and change the restored state (and we 
keep that). However, I also prefer "keep_fullscreen_state" so I maybe go 
with that.

>> +
>>                   TRACE( "restoring win %p/%lx\n", data->hwnd, data->whole_window );
>>                   release_win_data( data );
>>                   if ((style & (WS_MINIMIZE | WS_VISIBLE)) == (WS_MINIMIZE | WS_VISIBLE))
>>                       SetActiveWindow( hwnd );
>>                   SendMessageW( hwnd, WM_SYSCOMMAND, SC_RESTORE, 0 );
>> +                if ((data = get_win_data( hwnd )))
>> +                {
>> +                    data->keep_restored_state = FALSE;
>> +                    release_win_data( data );
>> +                }
> Also restrict it to only when restoring from minimized state. You can move release_win_data( data )
> to avoid a get_win_data() call.
> 

Sorry I don't get the part about restoring from minimized state. You 
mean checking for WS_MINIMIZE? We already check data->iconic which must 
be TRUE. And after all, that's the issue here (the X11 "minimized" 
state, not the Windows side). I don't think it's correct checking for 
WS_MINIMIZE at all.

It's true that SetActiveWindow causes the regression, but that's because 
d3d9's WM_ACTIVATEAPP resizes the window to fullscreen, which "saved" it 
by luck before. This bug, however, can manifest outside of d3d9 too, so 
it's not correct to only handle that case, even if that was the original 
motivation.

It's not a good idea to move release_win_data either, because those 
functions send messages and the app can potentially do crazy stuff in 
them. We don't want to stay in the critical section in that case.

However, to avoid entering the critical section when it's not needed, 
I'll add a check for NET_WM_STATE_FULLSCREEN the second time also.

> 
>>                   return;
>>               }
>>               TRACE( "not restoring win %p/%lx style %08x\n", data->hwnd, data->whole_window, style );
>> diff --git a/dlls/winex11.drv/window.c b/dlls/winex11.drv/window.c
>> index 4571739..d9d207c 100644
>> --- a/dlls/winex11.drv/window.c
>> +++ b/dlls/winex11.drv/window.c
>> @@ -1076,6 +1076,8 @@ void read_net_wm_states( Display* display, struct x11drv_win_data *data )
>>       if (!maximized_horz)
>>           new_state &= ~(1 << NET_WM_STATE_MAXIMIZED);
>>   
>> +    if ((data->net_wm_state & (1 << NET_WM_STATE_FULLSCREEN)) && data->iconic)
>> +        new_state |= 1 << NET_WM_STATE_FULLSCREEN;
> Please add some comments here.
> 

Noted.

>>       data->net_wm_state = new_state;
>>   }
>>   
>> @@ -2450,6 +2452,11 @@ void CDECL X11DRV_WindowPosChanged( HWND hwnd, HWND insert_after, UINT swp_flags
>>           if (event_type != ConfigureNotify && event_type != PropertyNotify &&
>>               event_type != GravityNotify && event_type != ReparentNotify)
>>               event_type = 0;  /* ignore other events */
>> +
>> +        /* if we keep the Wine window's restored state, we pretend it's
>> +           not an event, so that it's synced properly with the X server. */
>> +        if (data->keep_restored_state && !(new_style & WS_MINIMIZE))
>> +            event_type = 0;
> Is this really needed? Can this happen when handling ConfigureNotify?
> 

Yes. Consider a d3d9 application that hooks WM_ACTIVATEAPP and sends 
SetWindowPos from it (this is correct wrt to Windows, too). In this 
case, the window is WS_MINIMIZE while that happens, so we shouldn't 
reconfigure (sync) the X window, because on the Win32 side, SetWindowPos 
on a minimized window does nothing.

So even though 'keep_restored_state' is true, we still can't treat it as 
not an event, unless it's not minimized on the Windows side.

Because not being an event just reconfigures the X window to match Wine.

>>       }
>>   
>>       if (data->mapped && event_type != ReparentNotify)
>> @@ -2545,7 +2552,7 @@ UINT CDECL X11DRV_ShowWindow( HWND hwnd, INT cmd, RECT *rect, UINT swp )
>>           }
>>           goto done;
>>       }
>> -    if (!data->managed || !data->mapped || data->iconic) goto done;
>> +    if (!data->managed || !data->mapped || data->iconic || data->keep_restored_state) goto done;
> Same here.
> 

You mean adding a comment? Or whether it can happen? If this is the 
latter, this is obviously the most important part of the patch, because 
it *is* the root cause of the issue.

Without this check, ShowWindow overrides the state to the X server one, 
which is the non-fullscreen state (based on the spec) and that's 
obviously wrong and mismatches Windows.

The reason for the *previous* hook treating it as a non-event is 
because, if we *don't* override to the X state, then the Wine state 
(which is fullscreen for example) mismatches the X state. X server 
believes the window is restored to non-fullscreen since that's what it 
did, but Wine thinks it is fullscreen still. We don't sync it to be the 
X state, so we instead have to sync the X server to match Wine (i.e. we 
have to tell it to be fullscreen, instead).

And this sync (from Wine to X11) only happens when there's no event, 
since usually it's when Wine changes the window state via some API (not 
an event from outside Wine).

Thanks,
Gabriel