Repository navigation
renderer + embedded: the embedder presents a new host size with its bounds (set_host_present, host_frame) - #70
Merged
lawrencecchen merged 8 commits intoNov 3, 2026
Conversation
…ew-size frame first (red) cmux-next (frame owner refactor S4) changes a terminal view's bounds only in the Core Animation transaction that shows the frame drawn at the new size, or keeps both the old bounds and the old frame. Today a host-sized surface shows a new-size frame itself: setHostDrawableSize asks for a display pass, and the renderer sends the frame to main, where setSurfaceCallback takes any frame of the host size, whatever the view's bounds. The frame then lands in the old bounds in a later main-thread turn. These tests run on macOS with a real IOSurfaceLayer and real IOSurfaces (no GPU). Seams with no effect yet: Metal.setEmbedderPresents and IOSurfaceLayer.setEmbedderPresents (the mode), Metal.hostFrame (the embedder's query and present; returns .none) and apprt.HostFrame. With the embedder presenting sizes: - a new host size asks for no display pass, - the renderer does not show a frame of a new size; it keeps it as the fresh frame, - a main-thread present of a new size keeps the shown frame, - hostFrame reports READY, PRESENTED (the frame is the layer's contents in the caller's transaction), SHOWN, PENDING (counted late only when it presents), HELD and HIDDEN, - a frame that reaches the main thread late never replaces the frame the embedder showed (present sequence). They fail until the renderer leaves new sizes to the embedder. Controls pass now and must stay green: the default mode shows a new size as before and asks for a display pass, a first frame presents itself, a frame at the shown size still goes off main, a new-size frame that fits bounds the embedder already set is shown, and going back to the renderer shows a frame drawn for the host size.
…ounds (ghostty_surface_set_host_present, ghostty_surface_host_frame) cmux-next (frame owner refactor S4, ready-or-keep): for every terminal in every refresh, the view's bounds and the layer's pixels change together, or neither changes. With the renderer showing new sizes itself, a new-size frame went to main by dispatch_async and setSurfaceCallback took any frame of the host size, so it landed in the old bounds 2-4 refreshes after the geometry (cmux films P2). The waiting present (G1) cost main 4-9 ms per drag frame. Opt-in, macOS host-sized surfaces only; the default does not change. - ghostty_surface_set_host_present(surface, GHOSTTY_HOST_PRESENT_EMBEDDER): setHostDrawableSize asks for no display pass (a first frame still does); a display pass never waits (displayWhileEmbedderPresents); the renderer still draws each new size at once with cells rebuilt for it and keeps that frame (Metal.fresh, now with its present sequence), but setSurfaceCallback refuses a frame of another size than the shown one unless the layer's bounds already have its size. Frames at the shown size still go off main. Back to GHOSTTY_HOST_PRESENT_RENDERER, a frame drawn for the host size is shown at once. - ghostty_surface_host_frame(surface, present) -> ghostty_host_frame_e: NONE, HIDDEN, SHOWN, READY, PRESENTED, PENDING or HELD. With present, a READY frame becomes the layer's contents in the caller's transaction (IOSurfaceLayer.setSurfacePresented), ordered by the present sequence under the Core Animation lock, so a frame that reaches main later never replaces it. Never waits, never draws. Counted in host_present_calls, host_fresh_presents and host_late_presents. - publishFresh keeps only the newest frame by present sequence; a host size already on screen disarms publishing (no per-frame publish at rest). - ghostty.h parity tests for both enums (lib/enum.zig checkGhosttyHEnum). Chromium does the same for a resize: the browser compositor shows the new size only with a frame the renderer submitted for it, and keeps the old surface until then (fallback surface).
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the
 |
…mes, refused presents, the old present calls (red) Review of the embedder host present (62df0ad) found three gaps. These tests fail on 62df0ad: - an off-main frame at the presented size commits in its own transaction while the embedder's present (and the view's new bounds) is not committed yet, so a commit can carry the new pixels before the new bounds; - a present refused because the layer shows a newer frame leaves publishing disarmed, so no later frame of the host size can be presented; - ghostty_surface_present_host_frame(_ready) in the embedder mode still presents with setSurfaceSync (no present sequence), so a late frame replaces it.
…ead of an embedder present, a refused present re-arms, the old present calls follow it - IOSurfaceLayer.setSurfaceOffMain: while the layer shows the embedder's present (embedder_seq, set by setSurfacePresented), a completed frame does not commit in its own transaction, which could reach the screen before the embedder's transaction with the view's new bounds. It goes through main (setSurface), after that commit; the next frames go off main again. - Metal.presentFresh: a present refused because the layer shows a newer frame re-arms publishing, so the next frame of the host size can be presented. - Metal.presentHostSizedWait (ghostty_surface_present_host_frame and _ready) in the embedder mode is hostFrame(true): no wait, ordered by the present sequence. The default mode does not change.
…ps frames on main until the embedder committed (red) Second review of the embedder present (5a9907c): the off-main gate opened again at the first main-thread present after the embedder's. CFRunLoop runs main-queue blocks before the before-waiting observers, so a frame queued by the renderer can show in main's open transaction before the embedder's commit; the next frame then went off main and its own transaction could reach the screen ahead of the view's new bounds. The same after a frame shown because the bounds already fit. Seam with no effect yet: IOSurfaceLayer.setEmbedderCommitted (the embedder's 'my commit is done'). The existing gate test now reports the commit before it expects frames off main again.
…surface_host_frame_committed) Second review of the embedder present (5a9907c): the gate keyed on the embedder's present sequence opened again at the first main-thread present after it, which can run before the embedder's commit (a main-queue block), so the next frame could commit off main ahead of the view's new bounds. Ghostty cannot see the main thread's commit, so the embedder reports it: in GHOSTTY_HOST_PRESENT_EMBEDDER every main-thread contents change (a present, a frame shown on main, a sync present) marks the layer main-pending under the Core Animation lock, and setSurfaceOffMain refuses while it is set; ghostty_surface_host_frame_committed clears it. Without the call frames go through main (correct, slower). A mode change clears it. ghostty.h now says that the first frame's display pass and ghostty_surface_draw wait as they do by default, and that READY does not reserve the frame.
Review of 63c1398: - a frame of the shown size shown on main closed the gate, so terminal echo after a resize stayed on main until the next commit report (red); - the mode started with the gate open, though a change the main thread made before it may still be in its open transaction (red); - a sync present (ghostty_surface_draw) of a new size closes the gate (coverage, passes now) and one at the shown size keeps it open (red). The off-main control test and the bounds-fit test now report the embedder's commit first, since the mode starts closed.
…ts closed, and an off-main commit holds the CA lock Third review of the embedder present (63c1398): - An off-main transaction in GHOSTTY_HOST_PRESENT_EMBEDDER now takes [CATransaction lock] before begin and commits under it, and commits nothing when a check refuses. It can then never run inside the embedder's present pass (a layout or display pass of main's commit, which holds the lock), so it never ships a present or a frame main set after its checks, for this surface or another one in the window. ghostty.h now says the embedder presents and sets bounds under the lock or in such a pass. - Only a change of the shown size (a present, a first frame, a new-size frame shown on main or by a sync present) closes the gate; a frame of the shown size on main keeps echo off main. - The mode starts with the gate closed (a change main made before it may be in its open transaction); leaving it opens the gate. The default mode does not change.
Contributor
Author
|
Review (independent review subagent): SHIP at 9f88e48 Fresh independent review at the exact head (the 4th review of this PR; reviews 1-3 at 62df0ad, 5a9907c and 63c1398 each found bugs, and each was fixed with red tests first). Findings at 9f88e48, none blocking:
Runs (fork workflow, scratch branches never merged):
Earlier pairs: |
This was referenced Nov 3, 2026
lawrencecchen
added a commit
that referenced
this pull request
Nov 3, 2026
…lds the Core Animation lock renderer: an off-main present never commits while it holds the Core Animation lock (embedder-mode deadlock from #70)
This was referenced Nov 3, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
dispatch_async, andsetSurfaceCallbacktakes any frame of the host size whatever the view's bounds. In cmux's 120 Hz films the new-size frame landed in the old bounds 2-4 refreshes after the geometry, and two terminals landed 2 refreshes apart. The waiting present (ghostty_surface_present_host_frame, cmux G1) fixed the order but cost main 4-9 ms per drag frame.ghostty_surface_set_host_present(surface, GHOSTTY_HOST_PRESENT_EMBEDDER): a new host size asks for no display pass (a first frame still does), and a display pass never waits. The renderer still draws each new size at once with cells rebuilt for it and keeps that frame (Metal.fresh, now with its present sequence).setSurfaceCallbackrefuses a frame of another size than the shown one, unless the layer's bounds already have its size. Frames at the shown size still go off main. Back toGHOSTTY_HOST_PRESENT_RENDERER, a frame drawn for the host size shows at once.ghostty_surface_host_frame(surface, present)returnsghostty_host_frame_e:NONE,HIDDEN,SHOWN,READY,PRESENTED,PENDINGorHELD. Withpresent, a READY frame becomes the layer's contents in the caller's Core Animation transaction (IOSurfaceLayer.setSurfacePresented), ordered by the present sequence under the CA lock: a frame that reaches main later never replaces it. It never waits and never draws. It is counted inhost_present_calls,host_fresh_presentsandhost_late_presents.publishFreshkeeps only the newest frame by present sequence. A host size already on screen disarms publishing, so nothing publishes per frame at rest.release/cmux-9ce9b7d0-e2(cmux's pin34e73ba8b+ input: one compiled link regex per pattern for the whole process #69), so cmux can pin it without the v1.4.0 / Zig 0.17 merge onmain. A port tomainis a follow-up.Testing
30b8490a1, run https://github.com/manaflow-ai/ghostty-next/actions/runs/41611127503 on scratch branchscratch/host-present-red. That branch only narrows the test filters tohost present:.62df0adde, run https://github.com/manaflow-ai/ghostty-next/actions/runs/41611130888 onscratch/host-present-green. That branch addshost present:,ghostty.h HostPresentandghostty.h HostFrameto the full patch filter list.IOSurfaceLayerand real IOSurfaces, no GPU. Controls (the default mode, a first frame, an off-main frame at the shown size, a frame that fits bounds the embedder set, the switch back to the renderer) pass on both commits.ghostty_surface_present_host_frame(_ready)follow the ordered present in the embedder mode. Red: test commite999522d3, run https://github.com/manaflow-ai/ghostty-next/actions/runs/41615130156 (3 of 3 new tests fail; 92 other tests pass). Green: fix5a9907c45, run https://github.com/manaflow-ai/ghostty-next/actions/runs/41615133466 (ghostty-test 233/233, lib-vt 475/475, build, no-gettext and link smoke pass).TerminalHostFramePresenterwith this API, 120 Hz films, typing under load). That is the next cmux PR, after this one is pinned.Changelog
none