feat(capture): host arm of the fence ring #538

Merged

enricobuehler merged 3 commits from worktree-wp7-host-fence-v3 into main 2026-09-02 09:52:41 +00:00

Conversation 0 Commits 3 Files Changed 7

+461 -74

enricobuehler commented 2026-09-02 08:11:00 +00:00

Owner

Copy Link

Copy Source

Immunity plan WP7, step 3 of 3 — stacked on #535 (driver arm) ← #525 (protocol layer). Retarget down the stack as they merge. Replaces #537 (same tree plus the fence DACL fix, as one gate-clean commit; no force-push by house rule). With this, the fence protocol is live end to end wherever both sides open the fences; every other pairing keeps the keyed-mutex arm.

Negotiation (no flag day)

• Every ring generation carries two fresh shared ID3D11Fences (fresh, so both value sequences restart at zero on both sides; created with the SYSTEM-only DACL like every other frame object — design/idd-push-security.md item 10) in SetFrameChannelRequestV2, and a v4 header with the slot table. A pre-D3D11.4 device gets neither and stays on the mutex arm.
• The first ring on a box is a keyed-mutex probe carrying fences with CAP_FENCE_RING clear. The open-time attach's capability word says whether the driver opened them — remembered process-wide (DRIVER_FENCE_CAPABLE) — and a capable driver's probe ring is rebuilt on the fence protocol before any frame flows (one extra recreate on the first session per process). A fence-mode attach that fails falls back to the mutex arm and remembers that.
• Fence-mode textures are SHARED | NTHANDLE without a keyed mutex (HostSlot.mutex: Option); FrameChannelSender / send_frame_channel carry the v2 request (a pre-fence driver reads the v1 prefix).

Consume (D5)

Newest PUBLISHED slot above the last delivery from the slot table (fence::consumer_pick) — older publishes freed (newest-wins, the S2 finding), and any PUBLISHED record whose packed token names another generation freed too (a leftover of a superseded publisher; closes the race a ring rebuild opens). CAS PUBLISHED → READING, GPU Wait(ready, record.ready_value), the existing convert/blend path unchanged, Signal(retire, ++retire_value), publish the value in the record, then FREE (Release). A cursor regen re-reads the last slot only while it is FREE and retires it the same way. FenceSlotGuard puts a READING slot back to FREE on any early return.

Recreate

Fresh fences + slot table zeroed before the generation bump, CAP_FENCE_RING set/cleared per mode, v2 delivery.

Verification

• Writing gate passes. pf-capture / pf-vdisplay / punktfunk-host cannot build on macOS; windows-host dispatch on this branch is the compile proof. The driver arm's windows-drivers / driver-build already passed (run 22731).
• Live proof pending on the .173 box (both vendors, BGRA/FP16, chaos, process death); the S2 spike passed 10/10 there with the same shape, and WP4's attach log will show negotiated & CAP_FENCE_RING.

Immunity plan **WP7**, step 3 of 3 — stacked on #535 (driver arm) ← #525 (protocol layer). Retarget down the stack as they merge. Replaces #537 (same tree plus the fence DACL fix, as one gate-clean commit; no force-push by house rule). With this, the fence protocol is live end to end wherever both sides open the fences; every other pairing keeps the keyed-mutex arm. ## Negotiation (no flag day) - Every ring generation carries **two fresh shared `ID3D11Fence`s** (fresh, so both value sequences restart at zero on both sides; created with the SYSTEM-only DACL like every other frame object — `design/idd-push-security.md` item 10) in `SetFrameChannelRequestV2`, and a **v4 header** with the slot table. A pre-D3D11.4 device gets neither and stays on the mutex arm. - The **first ring on a box is a keyed-mutex probe** carrying fences with `CAP_FENCE_RING` clear. The open-time attach's capability word says whether the driver opened them — remembered process-wide (`DRIVER_FENCE_CAPABLE`) — and a capable driver's probe ring is **rebuilt on the fence protocol before any frame flows** (one extra recreate on the first session per process). A fence-mode attach that fails falls back to the mutex arm and remembers that. - Fence-mode textures are `SHARED | NTHANDLE` without a keyed mutex (`HostSlot.mutex: Option`); `FrameChannelSender` / `send_frame_channel` carry the v2 request (a pre-fence driver reads the v1 prefix). ## Consume (D5) Newest PUBLISHED slot above the last delivery from the slot table (`fence::consumer_pick`) — older publishes freed (newest-wins, the S2 finding), and **any PUBLISHED record whose packed token names another generation freed too** (a leftover of a superseded publisher; closes the race a ring rebuild opens). CAS PUBLISHED → READING, GPU `Wait(ready, record.ready_value)`, the existing convert/blend path unchanged, `Signal(retire, ++retire_value)`, publish the value in the record, then FREE (Release). A cursor regen re-reads the last slot only while it is FREE and retires it the same way. `FenceSlotGuard` puts a READING slot back to FREE on any early return. ## Recreate Fresh fences + slot table zeroed before the generation bump, `CAP_FENCE_RING` set/cleared per mode, v2 delivery. ## Verification - Writing gate passes. `pf-capture` / `pf-vdisplay` / `punktfunk-host` cannot build on macOS; **`windows-host` dispatch on this branch is the compile proof**. The driver arm's `windows-drivers / driver-build` already passed (run 22731). - **Live proof pending** on the `.173` box (both vendors, BGRA/FP16, chaos, process death); the S2 spike passed 10/10 there with the same shape, and WP4's attach log will show `negotiated & CAP_FENCE_RING`.

enricobuehler changed target branch from worktree-wp7-driver-fence-v2 to main 2026-09-02 09:13:49 +00:00

enricobuehler added 2 commits 2026-09-02 09:13:49 +00:00

feat(capture): host arm of the fence ring ... f5c0c5fd32

Immunity plan WP7, step 3 of 3. The IDD-push capturer runs the CAS plus
shared-fence slot protocol once the driver has proven it opens the
fences; every other pairing stays on the keyed-mutex arm.

Every ring generation carries two fresh shared ID3D11Fences (fresh so
both value sequences restart at zero on both sides), created with the
SYSTEM-only DACL and delivered in the v2 request, plus a v4 header with
the slot table. The first ring on a box is a keyed-mutex probe: the
attach's capability word says whether the driver opened the fences,
remembered process-wide; a capable driver's probe ring is rebuilt on the
fence protocol before any frame flows, and a failed fence-mode attach
falls back to the mutex arm.

Fence-mode consume: newest PUBLISHED slot above the last delivery from
the table (older publishes freed; a record whose packed token names
another generation freed too), CAS to READING, GPU-wait producer-ready,
convert as before, signal consumer-retire, publish the value, then FREE.
A RAII guard frees a slot on any early return. Fence textures are SHARED
plus NTHANDLE without a keyed mutex; the sender carries the v2 request.

fix(capture): open the attach wait to the fence-mode upgrade ... 6884833bb5

windows-host run 22795 on the host arm: wait_for_attach lives in the open child module as a private method, and the fence-mode upgrade after the probe ring calls it from the parent. pub(super), like create_ring_slots.

enricobuehler added 1 commit 2026-09-02 09:15:39 +00:00

Merge remote-tracking branch 'origin/main' into worktree-wp7-host-fence-v3 ... 754146b5c4

# Conflicts:
#	CHANGELOG.md

enricobuehler merged commit af11b908db into main 2026-09-02 09:52:41 +00:00

Sign in to join this conversation.

Reviewers

No Reviewers

Labels

No items

No labels

Milestone

No items

No Milestone

Projects

Clear projects

No projects

Assignees

Clear assignees

FernandoMX FlipperDan HitFrostbite Seenko bluelightspecial enricobuehler (Enrico Bühler) kszlim luxus nannku neoney sonar

No Assignees

1 Participants

Notifications

Subscribe

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: unom/punktfunk#538