-
Notifications
You must be signed in to change notification settings - Fork 8.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Maximizing/Restoring when the alt buffer is active crashes conhost #1206
Comments
|
This only repros if "Disable scroll forward" is enabled. This surprises me, considering the alt buffer doesn't need to do anything with _virtualBottom, but something's definitely busted there. |
This code looks bad:
if (gci.IsTerminalScrolling() && newViewport.Height() != _viewport.Height())
{
const short newTop = static_cast<short>(std::max(0, _virtualBottom - (newViewport.Height() - 1)));
newViewport = Viewport::FromDimensions(COORD({ newViewport.Left(), newTop }), newViewport.Dimensions());
}
During the resize down, we're resizing the alt buffer here. Prior to the
Looks like we're trying to set the top of the new viewport so that the virtual bottom stays anchored to the new viewport's bottom. However, this only makes sense when the viewport is getting bigger, not when it's getting smaller. In that scenario, we'll end up setting the top to somewhere in the middle of the buffer, and then be real confused by the next paint. |
## Summary of the Pull Request When a user had "Disable Scroll Forward" enabled and switched to the alt buffer and maximized the console, then restored down, we'd crash. Now we don't. ## References ## PR Checklist * [x] Closes #1206 * [x] I work here * [x] Tests added/passed ## Detailed Description of the Pull Request / Additional comments The problem is that we'd previously try to "anchor" the viewport to the virtual bottom when resizing like this. This would also cause us to move the top of the viewport down, into the buffer. However, if the alt buffer is getting smaller, we don't want to do this - if we anchor to the old _virtualBottom, the bottom of the viewport will actually be outside the current buffer. This could theoretically happen with the main buffer too, but it's much easier to repro with the alt buffer.
## Summary of the Pull Request When a user had "Disable Scroll Forward" enabled and switched to the alt buffer and maximized the console, then restored down, we'd crash. Now we don't. ## References ## PR Checklist * [x] Closes #1206 * [x] I work here * [x] Tests added/passed ## Detailed Description of the Pull Request / Additional comments The problem is that we'd previously try to "anchor" the viewport to the virtual bottom when resizing like this. This would also cause us to move the top of the viewport down, into the buffer. However, if the alt buffer is getting smaller, we don't want to do this - if we anchor to the old _virtualBottom, the bottom of the viewport will actually be outside the current buffer. This could theoretically happen with the main buffer too, but it's much easier to repro with the alt buffer. (cherry picked from commit c58033c)
🎉This issue was addressed in #2666, which has now been successfully released as Handy links: |
The "virtual bottom" marks the last line of the mutable viewport area, which is the part of the buffer that VT sequences can write to. This region should typically only move downwards as new lines are added to the buffer, but there were a number of cases where it was incorrectly being moved up, or moved down further than necessary. This PR attempts to fix that. There was an earlier, unsuccessful attempt to fix this in PR #9770 which was later reverted (issue #9872 was the reason it had to be reverted). PRs #2666, #2705, and #5317 were fixes for related virtual viewport problems, some of which have either been extended or superseded by this PR. `SetConsoleCursorPositionImpl` is one of the cases that actually does need to move the virtual viewport upwards sometimes, in particular when the cmd shell resets the buffer with a `CLS` command. But when this operation "snaps" the viewport to the location of the cursor, it needs to use the virtual viewport as the frame of reference. This was partially addressed by PR #2705, but that only applied in terminal-scrolling mode, so I've now applied that fix regardless of the mode. `SetViewportOrigin` takes a flag which determines whether it will also move the virtual bottom to match the visible viewport. In some case this is appropriate (`SetConsoleCursorPositionImpl` being one example), but in other cases (e.g. when panning the viewport downwards in the `AdjustCursorPosition` function), it should only be allowed to move downwards. We can't just not set the update flag in those cases, because that also determines whether or not the viewport would be clamped, and we don't want change that. So what I've done is limit `SetViewportOrigin` to only move the virtual bottom downwards, and added an explicit `UpdateBottom` call in those places that may also require upward movement. `ResizeWindow` in the `ConhostInternalGetSet` class has a similar problem to `SetConsoleCursorPositionImpl`, in that it's updating the viewport to account for the new size, but if that visible viewport is scrolled back or forward, it would end up placing the virtual viewport in the wrong place. So again the solution here was to use the virtual viewport as the frame of reference for the position. However, if the viewport is being shrunk, this can still result in the cursor falling below the bottom, so we need an additional check to adjust for that. This can't be applied in pty mode, though, because that would break the conpty resizing operation. `_InternalSetViewportSize` comes into play when resizing the window manually, and again the viewport after the resize can end up truncating the virtual bottom if not handled correctly. This was partially addressed in the original code by clamping the new viewport above the virtual bottom under certain conditions, and only in terminal scrolling mode. I've replaced that with a new algorithm which links the virtual bottom to the visible viewport bottom if the two intersect, but otherwise leaves it unchanged. This applies regardless of the scrolling mode. `ResizeWithReflow` is another sizing operation that can affect the virtual bottom. This occurs when a change of the window width requires the buffer to be reflowed, and we need to reposition the viewport in the newly generated buffer. Previously we were just setting the virtual bottom to align with the new visible viewport, but that could easily result in the buffer truncation if the visible viewport was scrolled back at the time. We now set the virtual bottom to the last non-space row, or the cursor row (whichever is larger). There'll be edge cases where this is probably not ideal, but it should still work reasonably well. `MakeCursorVisible` was another case where the virtual bottom was being updated (when requested with a flag) via a `SetViewportOrigin` call. When I checked all the places it was used, though, none of them actually required that behavior, and doing so could result in the virtual bottom being incorrectly positioned, even after `SetViewportOrigin` was limited to moving the virtual bottom downwards. So I've now made it so that `MakeCursorVisible` never updates the virtual bottom. `SelectAll` in the `Selection` class was a similar case. It was calling `SetViewportOrigin` with the `updateBottom` flag set when that really wasn't necessary and could result in the virtual bottom being incorrectly set. I've changed the flag to false now. ## Validation Steps Performed I've manually confirmed that the test cases in issue #9754 are working now, except for the one involving margins, which is bigger problem with `AdjustCursorPosition` which will need to be addressed separately. I've also double checked the test cases from several other virtual bottom issues (#1206, #1222, #5302, and #9872), and confirmed that they're still working correctly with these changes. And I've added a few screen buffer tests in which I've tried to cover as many of the problematic code paths as possible. Closes #9754
wsl
printf '\e[?1049h'
Notably, the crash doesn't occur for simple resizes -- even those covering more than one side.
The text was updated successfully, but these errors were encountered: