Terminal: BUG: Minimizing Terminal Crash

[evlli]

Environment

Windows build number: 10.0.18362.175
Windows Terminal version: 0.2.1831.0

Steps to reproduce

Execute a command that keeps outputting something for a while.
for example you can user apt list or ls in a directory with a lot of content.

Then, while the command is running, double click on the titlebar or click the minimize? button.

Expected behavior

The behavior I would have expected was for it to pause or keep running

Actual behavior

The Terminal Freezes and crashes after approximately 1,5 seconds
Other Terminal windows are unaffected

[zadjii-msft]

I'm not sure if it's the same crash, but while trying to repro this, I managed to get an exception while _restoring down_ from maximize while the app was outputting.

...
10 (Inline Function) --------`-------- conhost!wil::details::in1diag3::Throw_HrIf+0x17 [c:\ba\2009\s\dep\wil\include\wil\result_macros.h @ 5039] 
11 0000008b`002ff6d0 00007ff6`ceb4e2e5 conhost!TextBufferCellIterator::TextBufferCellIterator+0x281 [c:\ba\2009\s\src\buffer\out\textbuffercelliterator.cpp @ 43] 
12 (Inline Function) --------`-------- conhost!TextBuffer::GetCellDataAt+0x19 [c:\ba\2009\s\src\buffer\out\textbuffer.cpp @ 169] 
13 0000008b`002ff740 00007ff6`ceb4cefa conhost!Microsoft::Console::Render::Renderer::_PaintBufferOutput+0x1c5 [c:\ba\2009\s\src\renderer\base\renderer.cpp @ 573] 
14 0000008b`002ff8b0 00007ff6`ceb4ccca conhost!Microsoft::Console::Render::Renderer::_PaintFrameForEngine+0x1ea [c:\ba\2009\s\src\renderer\base\renderer.cpp @ 112] 
15 0000008b`002ff9a0 00007ff6`ceb4f6a2 conhost!Microsoft::Console::Render::Renderer::PaintFrame+0x6a [c:\ba\2009\s\src\renderer\base\renderer.cpp @ 65] 
16 (Inline Function) --------`-------- conhost!Microsoft::Console::Render::RenderThread::_ThreadProc+0x40 [c:\ba\2009\s\src\renderer\base\thread.cpp @ 165] 
17 0000008b`002ff9d0 00007ffd`26ef18a4 conhost!Microsoft::Console::Render::RenderThread::s_ThreadProc+0x52

Looked like it was trying to draw one of the lines that's now outside the viewport. Perhaps we're not locking the terminal properly for certain resize types? Or we've moved the viewport to a bad position on restore down? Fortunately, the repro is easy enough, so we just need to dig in more.

Thanks!

[zadjii-msft]

Leaving notes here for me after the weekend:

Trying to repro this, the terminal moves between [118x26]<->[207x61]

When we crash, the conpty's buffer has 61 rows with 118 chars each. We're trying to render row 60, and we want all 207 characters from it. Seems to imply that the buffer's width wasn't resized, but the height _was_.

That's not right.

[zadjii-msft]

Okay debugging more into this. For my repro, the resize is from [207x61] -> [118x26]

When we resize in ConPty, we call both SetConsoleScreenBufferInfoEx and if that succeeds, SetConsoleWindowInfo. In SetConsoleScreenBufferInfoEx, we'll perform _2_ ResizeWithReflow operations(!).

• The first will resize from [207x61] -> [118x61], changing the _width_ of all the rows.
• The second will resize from [118x61] -> [118x26], changing the height of the buffer. The second resize is triggered by the ScreenInfo::SetViewportSize call near the bottom ofSetConsoleScreenBufferInfoEx.
  
  
  
  • During this resize, we'll copy the first 26 rows (up to iOldRow = 25) just fine. However, the problem comes when we try to newline the cursor. When that happens on the 25th row, we'll attempt to circle the buffer, to shift all the buffer contents up a row, and give us a new row to work with. Hypothetically, this will happen 35 more times, as we copy all the old lines into the new buffer.
  
  
  • HOWEVER, as we call TextBuffer::IncrementCircularBuffer, we'll also tell the renderer that it should give the render engines a chance to paint, before any old text was lost. This is helpful when a newline is normally output to the terminal, but isn't really necessary to do during a resize.
  
  
  • The renderer will ask the VT engine if it wants to paint to handle the buffer circling. The VT engine will always say "yes".
  
  
  • The renderer will then _immediately_ render a new frame for the VT engine. Somewhere during this frame, the vt engine will throw an exception.
  
  
  • The exception will get caught way back up in SetConsoleScreenBufferInfoEx. The rest of the buffer will fail to get copied, and even worse, we _won't even move_ the [118x26] text buffer into the screen buffer, leaving the screen buffer with a incorrectly sized buffer.
  
  

Later on, we'll try and render a new frame, and try to render the 60th line of the buffer, with a width of 207 chars. The buffer _does_ incorrectly still have 61 lines, but it only has 118 chars per row, so the renderer throws an exception that's _not_ caught here, crashing conpty. Conpty crashing causes the terminal to crash.

Now we just need to come up with the right solution for handling this resize. Very simply, we could maybe tell conpty to not paint at all during a resize. Might be too risky a change for 0.3, however.

[mcpiroman]

@zadjii-msft Maybe, kind of, #1946? It got lost somewhere in the swarm of issues few weeks ago :(

[zadjii-msft]

@mcpiroman It's probably related to the larger class of "conpty resizing issues". I think it might be able to improve the QOL for the terminal by fixing this particular bug _now_, while a more complete design of the _right_ way to handle conpty resizing is a while off. I do still owe you a response on that other issue though.

[msftbot[bot]]

:tada:This issue was addressed in #2149, which has now been successfully released as Windows Terminal Preview v0.3.2142.0.:tada:

Handy links:

• Release Notes
• Store Download