feat(portal): follow the viewer's mouse mode with the embedded cursor (#203)

* feat(portal): follow the viewer's mouse mode with the embedded cursor

The portal stream hides the compositor cursor. A viewer that drives the pointer
with relative motion (UU's controlled-end mouse, also the Mac trackpad mode)
has no absolute position to draw its own pointer at, and the game's cursor
never reaches the picture, so the mouse can be moved and clicked but not seen.

Request cursor_mode=embedded while relative motion arrives, and keep it hidden
while absolute motion arrives: there the viewer draws its own pointer and an
embedded copy would trail it. The portal reads cursor_mode only when the
session is created, so switch by rebuilding the session with the saved consent,
the same path the privacy screen uses. A run has to last five events and 200 ms
before it counts, so a burst at connect or an alternating controller does not
rebuild; a failed rebuild backs off for five seconds.

The rebuilt session is checked against the monitor the previous one captured,
so a restored consent cannot silently move the share. The engine keeps the mode
for later opens (recovery, privacy, monitor switches) and skips the switch when
an older libuu-portal.so lacks uu_portal_set_cursor_mode.

UU_PORTAL_CURSOR_MODE=embedded sets the startup default.

* fix(portal): keep the cursor rebuild off locks and held input

Do not rebuild while the desktop is locked, and cancel an open that is already
waiting when it locks: GNOME closes the portal session on lock, so the request
would wait for consent that cannot be answered while holding `recovery`
against the post-unlock recovery.

Keys and buttons held through the portal input path are dropped when the
session is replaced, so wait for their release. Monitor selection takes the
same `recovery` lock, so it cannot install a stale source over the cursor
rebuild or the other way round, and the rebuilt session is revalidated against
the source it started from.

* fix(portal): carry the clipboard across cursor rebuilds

The engine keeps the viewer's text only in the session that provided it, and
the host re-sends it only for a new viewer update. A replacement session
therefore had no clipboard ownership, so changing the mouse mode made
previously synchronized text unpasteable until the viewer sent it again.

Offer the text to the replacement before the old session closes, but only
while that session still owns the selection: a local copy that has taken the
clipboard since is left alone. Touch consent upgrades replace the handle too,
so they take the same `recovery` lock as recovery, monitor selection and the
cursor rebuild and cannot install a handle opened under an older cursor mode.

* fix(portal): serialize clipboard writes with the cursor swap

The clipboard bridge calls clipboard_set from its own thread. Installing the
replacement handle and then carrying the old text over left a window where an
update could land on the new session and be overwritten by the older text,
with no pending update left to restore it.

Take a clipboard write lock before the handle is installed and hold it until
the carry-over finishes; clipboard_set takes the same lock. A write that
starts before the swap is included in the text carried over, a write that
starts after waits and wins.

* fix(portal): cancel a waiting rebuild when the desktop ends

The desktop ends its share when the last controller leaves (end_idle), but an
open that is still waiting for consent is not the installed handle, so it
survived that shutdown and installed a live share afterwards.

Bump an end generation in Session::end(), snapshot it before the open, and
cancel the request when it moves. The generation is checked again under the
handle lock before installing, so an end that races the swap either aborts
the rebuild or ends the handle it just installed.

* fix(portal): keep session swaps consistent with the viewer's state

Addresses the ownership race reported on the previous commit and the paths
that share its failure mode.

uu_portal_clipboard_set publishes the new text before SetSelection so a
transfer during the call serves it, restores the old text when the call fails,
and ignores a stale revision so a newer owner wins. The engine reports the
cursor mode it actually requested and the modes the backend advertised; the
host follows that instead of assuming the request was applied and stops
retrying where an embedded cursor is unavailable. The end generation is
sampled before the rebuild queues for `recovery`, a rebuild refuses to start
while the session is no longer alive, and cancellation now covers virtual
output opens.

Portal-bound keys and buttons are tracked separately, so an input held through
the portal defers the swap on every routing path, and the check is repeated
after the open. Every handle replacement (privacy screen, monitor switch,
touch consent, recovery after lock or stream loss) carries the viewer's
clipboard under the clipboard write lock, and the rebuilt session is checked
against the captured geometry.

* fix(portal): stop the swap bookkeeping from outliving a deferred rebuild

open_inner publishes the replacement's cursor mode before the swap. When held
input defers the rebuild, that session is never installed, so the host belief
described the discarded handle and later motion looked like it already matched.
Restore the belief from the installed session on every abort path.

The extra portal releases in release_held now stop at the first failure like
the sweep above; with a dead portal bus each remaining input waited out the
three-second timeout while the input and desktop-owner locks were held.

* fix(portal): invalidate clipboard ownership when a session ends

A closed session stops receiving owner changes, so a cached clipboard_owns
survived into the recovery paths and could republish old viewer text over a
newer local copy. Clear the claim in uu_portal_end and the Closed handler, and
let uu_portal_clipboard_copy refuse a source that is no longer alive.

Mark the embedded cursor unavailable when the engine did not apply it: with
AvailableCursorModes unreadable the session opens hidden, and retrying on every
motion run rebuilt it every 200 ms without changing the mode.

* fix(portal): retire held portal input when the devices are replaced

A key or button held through the native portal path is recorded when its
injection succeeds. If the session dies first, the key-up fails and the record
survives; recovery then installs new devices while the tracker still says an
input is held, so pointer-mode detection keeps deferring.

Keep the tracker in module state and clear it whenever the portal replaces or
ends its devices: the rebuild, both recovery paths and Session::end.

* fix(portal): scope held-input retirement to the retired handle

A press delivered to the replacement after the swap could be recorded while
the old handle was still being retired, and the unconditional reset then
deleted it. Hold the recording lock from before the swap until the reset
finishes, so only inputs sent to the retired devices are dropped.

A portal touchscreen contact was not tracked at all: the swap retired the
active contact while TouchDrag stayed down, so the following move reached a
session that never saw the press. Track contact down/up and include it in the
deferral state.

* fix(portal): serialize touch injection with handle replacement

inject_touch wrote the contact state outside the recording lock the handle
swap holds, so a contact delivered to the replacement between the swap and the
reset could be cleared, and the following phase-1 would fail on new devices
that never saw the press. Take the same lock across the injection and the
state update.

* fix(portal): decide the swap under the held-input lock

The final held-input check ran before install_replacement() took the serial
lock, so an input could be injected and recorded against the old handle while
the install waited for the clipboard lock. The install then retired the
records and swapped, destroying a contact the new devices never saw. Repeat
the check under the serial lock that recording uses, and defer the rebuild
when it hits.

* fix(portal): hold the input lock across native injection, refresh privacy geometry

inject_button and inject_key injected through the portal and only took the
recording lock afterwards, so a swap could take the lock, see no held input
and replace the devices that had just received the press. Take the serial lock
across the injection and its record, as the touch path does; the record
helpers no longer take it themselves.

A cursor rebuild that replaces the privacy virtual output invalidates the
cached pointer region: recompute it from the settled layout so absolute input
keeps landing inside the captured share.
This commit is contained in:
Mr. Error authored and GitHub committed 2026-11-09 08:36:37 +08:00
1 parent c489f3056f
commit 21e9e690ae
11 files changed
+592 -38

No files matched your search

+1
View File
@@ -357,6 +357,7 @@ systemctl --user restart uuremote-native.service
| `UU_X11_OUTPUT` | X11 默认采集的显示器名称(`xrandr --listmonitors`),默认主显示器 |
| `UU_WAYLAND_NO_VIRTUAL_KEYBOARD=1` | 原生 Wayland 不创建虚拟键盘(如 Denial 只把它给自己的输入法),键盘改走 uinput。合成器没有 wlr-virtual-pointer 时鼠标也自动改走 uinput(#72) |
| `UU_INPUT_BACKEND` | `auto`(默认,Wayland 下可用时用 uinput)/ `uinput` / `native` |
| `UU_PORTAL_CURSOR_MODE` | 启动时的光标模式:`embedded` 让合成器把指针画进抓屏流,默认隐藏。运行中按客户端发来的相对/绝对位移自动切换;每次切换重建一次 portal 会话,画面和输入短暂中断 |
| `UU_TERMINAL_SHORTCUTS=0` | 关闭终端窗口里的 Ctrl+V 粘贴转换 |
| `UU_VIDEO_ENCODER` | `cpu` 强制 CPU;`vaapi` 显式启用 Intel/AMD 硬编;`nvenc` 强制 NVENC;未设置或 `auto` 时后台探测(每个候选编码一帧,限时 3 秒)。有 NVIDIA 驱动时顺序是 NVENC → CPU;没有 NVIDIA 时只有 CPU,不探测。探测未完成就建流则沿用旧规则(有 NVIDIA 用 NVENC,否则 CPU)。自行编译的 `--features inproc-av` 上,仅 X11 会先走进程内引擎:外部 FFmpeg 探测失败时记为 `native_unprobed`,沿用旧规则(NVENC);Wayland、portal、KMS 回退 CPU。只有 CPU 时跳过探测,记为 `not probed (only candidate)`。探测成功和这种只含 CPU 的跳过会缓存到 NVIDIA 是否存在发生变化为止(`/proc/driver/nvidia/version` 或 `/dev/nvidia0`)。之后的桌面连接在后台再探测一次(不挡住这一路,仍用缓存的编码器)。NVENC 探测失败、选用 CPU 时只缓存 10 分钟;到期后下一次桌面连接同样在后台再探测一次,再次失败则重新计 10 分钟。显式 `cpu` / `vaapi` / `nvenc` 不启动探测。`native_unprobed` 不算失败缓存。优先级低于 `host_settings.json` 里非 `auto` 的 `video_encoder`。VAAPI 要等 #33 解决后才会进入 auto |
| `UU_VAAPI_BUFFER_MS` | VAAPI 的 VBV 缓冲(10–1000 ms,仅 `UU_VIDEO_ENCODER=vaapi` 时生效);高分辨率下关键帧被压糊时用它换稳定画质 |