[PATCH setup] Make sure that fatal error messages are visible

[Ken Brown]

On 12/20/2017 12:09 PM, Ken Brown wrote:
> On 12/20/2017 11:19 AM, Jon Turney wrote:
>> On 19/12/2017 00:53, Ken Brown wrote:
>>> The message box produced by TOPLEVEL_CATCH could be hidden by whatever
>>> window was previously being displayed, so that setup appeared to hang.
>>> Fix this by giving fatal error message boxes type MB_SETFOREGROUND.
>>
>> This is good as far as it goes, but is kind of working around the fact 
>> that fatal() is being called with an NULL owner HWND.
>>
>> This is not idea because I guess it means that propsheet window is 
>> still activate-able when this messagebox is displayed (MB_APPMODAL 
>> doesn't apply)?
> 
> It turns out that MB_SYSTEMMODAL did the job.  I tried MB_APPLMODAL and 
> MB_TASKMODAL also, but both of those still allowed me to activate the 
> propsheet window.

I just discovered that MB_SYSTEMMODAL doesn't prevent activation of the 
propsheet in all circumstances.  I tried the following:  I created a 
corrupt local package in the local cache and then tried to reinstall the 
package.  This led to an exception thrown by check_for_cached(), which 
normally would be caught by do_download_thread().  But I disabled that 
as follows:

--- a/download.cc
+++ b/download.cc
@@ -278,8 +278,8 @@ do_download_thread (HINSTANCE h, HWND owner)
           catch (Exception * e)
             {
               // We know what to do with these..
-             if (e->errNo() == APPERR_CORRUPT_PACKAGE)
-               fatal (owner, IDS_CORRUPT_PACKAGE, pkg.name.c_str());
+             // if (e->errNo() == APPERR_CORRUPT_PACKAGE)
+             //        fatal (owner, IDS_CORRUPT_PACKAGE, 
pkg.name.c_str());
               // Unexpected exception.
               throw e;
             }

As I result the exception was caught by the TOPLEVEL_CATCH for the 
download thread.  When the fatal message box was shown, I was still able 
to interact with the download progress reporting window.  Fortunately, 
the only available button there was 'Cancel', so no harm was done.

In spite of this, MB_SYSTEMMODAL is still the best option I've been able 
to come up with.

Ken