Repository navigation
End capture before removing the headless output on every exit, and cap headless resizes - #64
Merged
dhh merged 4 commits intoOct 24, 2026
Conversation
The screen session cleared its dmabuf and wl_shm format lists after every constraints done. When the first GPU encode failed, UseShm dropped GBM and reallocated the ring against those empty lists, so the fallback failed with "no usable wl_shm format offered (got [])" and the session ended instead of capturing into wl_shm. Constraint events now fill a pending set, and done promotes it to a complete set and starts the next batch empty, since the compositor sends the full set in every batch. All allocation reads only the complete set, so a reallocation between batches sees the last full set and never pairs its formats with the size of a batch still in flight. The cursor session goes through the same struct and now also sizes its buffer from the complete set. --capture dmabuf still refuses the fallback. The per-event constraint handling moves into Constraints::apply, and the ring format choice into dmabuf_format and shm_choice, so the unit tests drive the same code as the session handler and allocate_ring.
The ordered teardown at the end of the session loop (release held keys, end the capture thread, end the input thread, then remove the headless output) ran only when the loop ended normally. A `?` return from a client message (a second Hello), a writer error (the ssh link dropping) or the encoder, or a panic, dropped Session in field order instead. `output` is declared before `input` and `capturer`, so the headless output was removed while the cursor capture session was still alive, which aborts Hyprland 0.56 (docs/hardware-quirks.md). Hold the output with the capture and input threads in ServedOutput, a small generic type whose Drop ends both threads before the output drops. Session derefs through it to the output, so the call sites read as before, and its own Drop releases held keys first, which runs before any field drops. The normal exit now just drops the session. A test drops a ServedOutput of stand-ins normally and under a panic and checks the order.
Resize clamped a headless output only to max_width and max_height from the client's own Hello, so a client could ask Hyprland for a 16384x16384 output and the capture ring would then allocate four buffers that size. The 1920x1080 clamp in setup_output applies only to the first mode. Cap a resize at 4096 a side and 3840x2160 worth of pixels before clean_mode, scaling both sides by one factor so the aspect ratio holds. The gliff client never asks for more than 3840x2160, and a larger output would only be scaled down to the encoder maximum for the stream. The capture thread follows size changes on its own, so CaptureConfig gains max_size, a SizeLimit with a side and a pixel count, and allocate_ring refuses a ring past it with an error. The server sets it, from the same constant as the resize cap, for headless outputs only: a mirrored monitor is the size it is. The resize size and the capture config are built in headless_target and capture_config, which resize and start_capture call and the tests exercise.
 [Medium impact] The PR appears safe to merge; no actionable issues were found. SummaryThe PR makes session cleanup run on normal exits, errors, and unwinding panics.
|
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.
Headless sessions had two gaps.
?return or an unwinding panic droppedSessionin field order instead. Such a return comes from a second Hello, a writer error when the ssh link drops, or an encoder failure. That order removed the output while the cursor capture session was still alive, which docs/hardware-quirks.md records as aborting Hyprland 0.56.2. The output and both threads now live inServedOutput, whose Drop ends the threads before the output on every exit. The session's own Drop releases held keys first.Resizewas bounded only by the size the client declared in its own Hello, so a client could request a 16384x16384 output and a four-buffer capture ring of that size. Resize is now capped at 4096 a side and 3840x2160 worth of pixels, keeping the aspect ratio. The capture thread refuses a headless ring past the same limit, since it follows size changes on its own. Mirrored outputs are unchanged.Validation:
ServedOutput(normal drop and panic), the resize cap and its aspect ratio, and the capture limit. Each fails with the code it covers removed.scripts/e2e.sh(nested Hyprland, AMD Radeon 840M): the headless server and client steps, the probe, pipeline and CPU tier steps pass. The run then stopped at step 6, the e2e race that Wait for the client's clipboard item in e2e, not a MIME type #65 fixes.scripts/check.shpasses.This conflicts in
crates/gliff-server/src/session.rswith #66 and #73, and with #47, and merges cleanly with #57. I am happy to rebase on whichever lands first.Opus 5.5 high, on behalf of TechLuddite.
Implemented with Claude Opus 5.5 at high effort and reviewed by GPT-6.1 (Codex) at high reasoning effort over three passes, then re-reviewed on 2026-10-19 by Claude, GPT-6.1 and Grok 4.7 at medium effort.