The RGB→NV12 compute CSC shares the SMs with the game; request an
elevated global-priority queue (VK_KHR/EXT_global_priority) so the
driver can schedule it ahead. PUNKTFUNK_VK_QUEUE_PRIORITY =
off|high|realtime (default realtime), downgrading REALTIME→HIGH→none
on NOT_PERMITTED/INITIALIZATION_FAILED — a refused class never fails
the bridge. Mirrors PyroWave's ac0e7332 (same honest caveat: WDDM
didn't move; the Linux driver may — contended-tail lever, correct and
harmless either way).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Bind the session's IO streams to the encode thread's high-priority copy
stream: in sync-retrieve depth-1 use the per-frame input copy and cursor
blend now enqueue with NO cuStreamSynchronize and encode_picture orders
after them on the stream. Same stream both directions, so the encode's
completion is inserted into the stream and later work (the next frame's
copy into a reused ring slot) waits for it.
Soundness gate: the fast path engages only when pending is empty (true
depth-1 usage) — every prior encode was drained by a blocking poll, and
the caller holds the frame payload across the matching poll (contract now
documented on Encoder::submit; both host loops already comply). Pipelined
callers and PUNKTFUNK_NVENC_ASYNC mode keep the blocking copies.
True zero-copy input registration (registering the worker-owned IPC
buffer directly) stays the LN2 v2 follow-up — it needs a contiguous
worker-pool NV12 layout and a registration<->IPC-mapping lifetime tie.
PUNKTFUNK_NVENC_STREAM_ORDERED=0 restores the old blocking behavior.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Seven medium findings from the round-1 sweep, all in pf-zerocopy. Adjudicated
against source; all zero-copy-preserving (no GPU→CPU→GPU roundtrip introduced).
- import_src (vulkan.rs) leaked a VkBuffer + VkDeviceMemory + dup'd fd on every
fallible step before the success-only src_cache.insert. A failed import is
survived by the worker and RETRIED by the caller every frame, so this leaked
per frame for the worker's whole lifetime. Now owns the dup fd via OwnedFd
(closes on early return until Vulkan consumes it at allocate_memory) and
destroys the buffer + frees memory on each error path.
- ensure_dst destroyed the old exportable buffer and nulled self.dst BEFORE
building the replacement, so a failed rebuild both dropped the working buffer
and leaked every object the partial rebuild created (raw ash handles, no Drop;
VkBridge::drop only frees the live self.dst). Now builds fully, unwinds each
error locally, and swaps only on success.
- The submit→wait→reset sequence in import_linear_nv12 / import_linear `?`-ed out
of wait_for_fences BEFORE reset_fences on a TIMEOUT/DEVICE_LOST, leaving the
shared self.cmd PENDING and self.fence IN-USE while the caller retries on the
same bridge (and ensure_dst later destroys dst.buffer assuming nothing is in
flight). Now drains the GPU (device_wait_idle) and resets the fence before
propagating — fixing the reported ensure_dst UAF at its source.
- copy_pitched_nv12_to_buffer writes height.div_ceil(2) chroma rows into a UV
plane sized at height/2, so an odd height overruns by one uv_pitch row (OOB
device write / CUDA_ERROR_ILLEGAL_ADDRESS). Added the even-dimension guard to
import_linear_nv12, matching Nv12Blit/Yuv444Blit.
- sweep_reaper only reaped already-exited workers, so a worker wedged inside a
driver call lingered forever holding its CUcontext + BufferPool (hundreds of MB
VRAM). Now force-kills a child parked past REAPER_KILL_DEADLINE (20s).
Compile + tests green on Linux .21 (RTX 5070 Ti): pf-zerocopy 17/0. The error
paths themselves are not fault-injected; the fixes are structural.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The three high-severity defects from the round-1 sweep of pf-inject /
pf-zerocopy / pf-capture (adjudicated against source — all 7 reported criticals
in these crates downgraded; these were the real highs).
- pf-inject steam_gadget: `SteamDeckGadget::drop` set `running=false` then joined
the control thread, which spends steady state parked in a blocking, no-timeout
`EVENT_FETCH` ioctl that only tests `running` at its loop top. The flag never
reaches it, closing the fd can't wake an in-flight ioctl (the syscall holds a
file reference, and the fd is shared via Arc by the very threads being joined),
so the join hung — and it runs on the session input thread via
`PadSlots::sweep`, driven by the client's `active_mask`, so a remote peer
clearing its pad bit could freeze all session input. Now wakes the parked
threads with SIGUSR1 (no-op, non-SA_RESTART handler → the ioctl returns EINTR
and the loop exits), retried until each reports done and bounded (~1s).
- pf-zerocopy cuda: the GL→CUDA "sync point" was never established for the copy.
`cuGraphicsMapResources`/`UnmapResources` were issued on the NULL stream, but
the D2D copy runs on `copy_stream()`, a `CU_STREAM_NON_BLOCKING` stream exempt
from implicit NULL-stream ordering — and the GL de-tile/CSC that produced the
texture ends with only `glFlush` (no fence). So the copy could race ahead of
the not-yet-retired GL draw: intermittent stale/torn frames under GPU load, on
the default NVIDIA capture→encode path. Map, copy, and unmap now share
`copy_stream()`, so map's device-side guarantee orders the GL work before the
copy. Zero-copy preserved (no GPU→CPU→GPU roundtrip).
- pf-capture cursor meta: `update_cursor_meta` trusted three producer-written
fields (bitmap_offset, pixel offset, stride) with no bound against the metadata
region, driving OOB pointer arithmetic and an oversized `from_raw_parts` — an
OOB read that SIGSEGVs inside the PipeWire `.process` callback (uncatchable by
the surrounding `catch_unwind`). Switched to `spa_buffer_find_meta` to obtain
the region's real `size` and validate every offset with checked arithmetic
before each deref/slice, mirroring the fd-length guard the main frame path
already applies.
Compile + existing tests green on Linux .21 (real RTX 5070 Ti): pf-inject 74/0,
pf-zerocopy 17/0, pf-capture 1/0. The gadget deadlock path only executes on a
SteamOS host with raw_gadget/dummy_hcd (not reproducible on the CachyOS box), so
that fix is reasoned + compile-verified, not runtime-exercised.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
design/latency-reduction-2026-07.md T2.5's Linux half: the LINEAR dmabuf path
(gamescope's only offer) fed NVENC RGB, paying its internal RGB->YUV CSC on
the SM the game is saturating — the exact contention §5.A removed everywhere
else. The Vulkan bridge now carries a buffer-to-buffer RGB->NV12 compute
shader (rgb2nv12_buf.comp, BT.709 limited, coefficient-identical to
pf-encode's rgb2yuv.comp; whole-word writes so no 8-bit-storage feature is
needed): import dmabuf -> dispatch CSC into the exportable buffer -> CUDA
de-strides both planes into a pooled two-plane NV12 buffer. PUNKTFUNK_NV12
(default-on) now covers LINEAR; a CSC failure latches RGB for the stream
(mid-frame fallback, no dropped frame); 4:4:4 LINEAR sessions stay RGB (never
silently subsample). New ImportKind::LinearNv12 rides the existing worker IPC
(appended last per the wire-tag rule); cursor stays downstream (blend_nv12).
Validated: .21 clippy -D warnings (pf-zerocopy/pf-capture/host+nvenc) + 17
zero-copy tests. Owed: on-glass gamescope session (visual + dmon sm% check).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The eight W6 leaf crates hardcoded 0.12.0 instead of inheriting the
workspace version — switched to version.workspace = true so the next bump
is one line again.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
linux/zerocopy/* (CUDA context/buffers + EGL/Vulkan dmabuf import + the isolated
import worker) and linux/dmabuf_fence.rs move wholesale into crates/pf-zerocopy,
so the coming pf-frame vocabulary crate (FramePayload::Cuda owns a DeviceBuffer)
and the pf-encode/pf-capture subsystem crates can reach the GPU plumbing without
the host orchestrator in between (plan §W6). Content stays Linux-only; the crate
compiles to an empty lib elsewhere, so dependents carry a plain dependency.
drm_fourcc deliberately does NOT move: it consumes the frame vocabulary
(PixelFormat), which sits ABOVE pf-zerocopy — it lives with capture for now and
moves into pf-frame next. cuda's ffi re-export bumps pub(crate)->pub (the raw
CUdeviceptr vocabulary is consumed across the crate boundary by the encode
backends). A crate::zerocopy shim module keeps every existing path valid until
capture/encode themselves move out.
Verified: Linux clippy -D warnings (pf-zerocopy --all-targets + host
nvenc,vulkan-encode,pyrowave --all-targets) + 17/17 pf-zerocopy tests + 321/321
host tests; Windows clippy nvenc,amf-qsv --all-targets Finished exit 0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
enricobuehlerandClaude Fable 5