[MG_Backend, MG_Remote] (P7): draft B4 aggregate waits and staged extent reset

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.cpp

@@ -307,8 +307,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (result != VK_SUCCESS) {
             return result;
         }
-        // The slot's fence has been waited: every command buffer this slot
+        // Under disaggregation VulkanRenderer waits every registered fence
-        // submitted (including mid-frame flushes) has finished executing.
+        // through this slot's last submit before calling here. That aggregate
+        // proof covers mid-frame pooled-fence flushes too; the slot fence alone
+        // would not. Initialize's first acquire has no prior submissions.
         FreeRetiredCommandBuffers(frame);
 
         result = vkAcquireNextImageKHR(device, swapchain, timeout, frame.imageAvailableSemaphore, acquireFence,

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.cpp

@@ -808,8 +808,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Nothing bumps the slice epoch for those other resources, so freeing the buffer
         // here left the streamed memo handing a destroyed VkBuffer to vkCmdBindIndexBuffer
         // (llvmpipe then faulted inside the draw; the Create/Flywheel indirect retrace died
-        // exactly this way). Mid-frame drains do not advance m_frameSerial, so they must not
+        // exactly this way). Most mid-frame drains do not advance m_frameSerial; every
-        // free arena storage either: the arena's own ResetFrame/BeginFrame is the point where
+        // eighth does, but only through BeginFrame after the arena/memo boundary work.
+        // The drain's per-resource sweep must not free arena storage: ResetFrame/BeginFrame is where
         // the slot's slices stop being reachable, and that is where these releases land.
         for (Uint32 frameIndex = 0; frameIndex < m_deferredBufferReleases.size(); ++frameIndex) {
             CollectDeferredReleases(frameIndex);

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -7063,6 +7063,10 @@ void main() {
 #if MOBILEGL_BUILD_DISAGGREGATED
         if (MG_Config::Transport != MG_Config::TransportMode::Monolith) {
             RewindWireDescriptorSetsIfDue();
+            // The wire route returns before the monolith branch's draw-gated
+            // sweep. Dead wire texture/renderbuffer records otherwise live
+            // until the next frame boundary, which a long trace may not reach.
+            m_textureManager->CollectGarbage();
             return SetupWireDraw(frame, mode, aspects, drawParams, pIndexBufferView);
         }
 #endif
@@ -7685,6 +7689,7 @@ void main() {
 #if MOBILEGL_BUILD_DISAGGREGATED
         if (MG_Config::Transport != MG_Config::TransportMode::Monolith) {
             RewindWireDescriptorSetsIfDue();
+            m_textureManager->CollectGarbage();
             DispatchWireCompute(numGroupsX, numGroupsY, numGroupsZ);
             return;
         }
@@ -13527,12 +13532,32 @@ void main() {
             return false;
         }
         if (m_device == VK_NULL_HANDLE || m_graphicsQueue == VK_NULL_HANDLE) {
+#if MOBILEGL_BUILD_DISAGGREGATED
+            return IsFrameSerialComplete(serial);
+#else
             return true;
+#endif
         }
-        // Every submission is recorded with the frame serial it was made under, so the wait can be
+#if MOBILEGL_BUILD_DISAGGREGATED
-        // narrowed to the first submission at or past the requested serial instead of draining the
+        // More than one submit can carry the same frame serial (mid-frame flush,
-        // whole queue. OnSubmitsCompletedUpTo calls NotifyFrameSerialComplete for every record it
+        // then Present). The completed floor is clamped below any serial still
-        // retires, so the completed-serial floor still advances correctly after one fence wait.
+        // held by an in-flight record; the first matching fence is not proof.
+        // OnSubmitsCompletedUpTo erases records, so copy fields and restart the
+        // search after every wait instead of continuing an invalidated iterator.
+        for (;;) {
+            const auto record = std::find_if(m_inFlightSubmits.begin(), m_inFlightSubmits.end(),
+                [serial](const auto& candidate) {
+                    return candidate.frameSerial >= serial && candidate.fence != VK_NULL_HANDLE;
+                });
+            if (record == m_inFlightSubmits.end()) break;
+            const Uint64 submitIndex = record->submitIndex;
+            if (!WaitForSubmitsUpTo(submitIndex, UINT64_MAX))
+                break; // fall through to the queue drain below
+            TryDrainFrameTransients();
+            if (IsFrameSerialComplete(serial)) return true;
+        }
+#else
+        // Pull build's historical one-fence wait stays byte-identical.
         for (const auto& record : m_inFlightSubmits) {
             if (record.frameSerial < serial || record.fence == VK_NULL_HANDLE) {
                 continue;
@@ -13546,6 +13571,7 @@ void main() {
             TryDrainFrameTransients();
             return true;
         }
+#endif
 
         // No usable record - fall back to draining the graphics queue. This over-waits (bounded by
         // the in-flight frame count) but never deadlocks.
@@ -13559,7 +13585,11 @@ void main() {
         // The queue was just drained; take the free frame-boundary drain when
         // nothing is recorded (present-less timer-query loops). No-op otherwise.
         TryDrainFrameTransients();
+#if MOBILEGL_BUILD_DISAGGREGATED
+        return IsFrameSerialComplete(serial);
+#else
         return true;
+#endif
     }
 
     Uint64 VulkanRenderer::GetSyncPointSubmitIndex() const {
@@ -13573,7 +13603,12 @@ void main() {
             return false;
         }
         const auto& frame = m_frameContext.GetCurrent();
+#if MOBILEGL_BUILD_DISAGGREGATED
+        return frame.isCommandRecording || frame.hasCommandBufferRecorded ||
+               frame.isPreCommandRecording || frame.hasPreCommandBufferRecorded;
+#else
         return frame.isCommandRecording || frame.hasCommandBufferRecorded;
+#endif
     }
 
 #if MOBILEGL_BUILD_DISAGGREGATED
@@ -13639,6 +13674,29 @@ void main() {
         }
     }
 
+#if MOBILEGL_BUILD_DISAGGREGATED
+    Bool VulkanRenderer::WaitForSubmitsUpTo(Uint64 submitIndex, Uint64 timeoutNs) {
+        if (submitIndex <= m_completedSubmitCounter) return true;
+        if (submitIndex > m_submitCounter || m_device == VK_NULL_HANDLE) return false;
+        Vector<VkFence> fences;
+        for (const auto& record : m_inFlightSubmits) {
+            if (record.submitIndex > submitIndex) break;
+            if (record.fence == VK_NULL_HANDLE) return false;
+            fences.push_back(record.fence);
+        }
+        if (fences.empty()) return false;
+        const VkResult result = vkWaitForFences(m_device, static_cast<Uint32>(fences.size()),
+                                                fences.data(), VK_TRUE, timeoutNs);
+        if (result == VK_SUCCESS) {
+            OnSubmitsCompletedUpTo(submitIndex);
+            return true;
+        }
+        if (result != VK_TIMEOUT)
+            MGLOG_E_ONCE("WaitForSubmitsUpTo: vkWaitForFences returned %d", result);
+        return false;
+    }
+#endif
+
     void VulkanRenderer::OnSubmitsCompletedUpTo(Uint64 submitIndex) {
         m_completedSubmitCounter = std::max(m_completedSubmitCounter, submitIndex);
 #if MOBILEGL_BUILD_DISAGGREGATED
@@ -13945,7 +14003,12 @@ void main() {
         // pooled fences and mid-frame command buffers.
         RefreshCompletedSubmits();
         auto& frame = m_frameContext.GetCurrent();
+#if MOBILEGL_BUILD_DISAGGREGATED
+        if (!frame.isCommandRecording && !frame.hasCommandBufferRecorded &&
+            !frame.isPreCommandRecording && !frame.hasPreCommandBufferRecorded) {
+#else
         if (!frame.isCommandRecording && !frame.hasCommandBufferRecorded) {
+#endif
             // GL flush semantics still demand batched texture uploads start
             // executing in finite time even when no draw was recorded.
             if (m_textureManager) {
@@ -13988,9 +14051,15 @@ void main() {
         const VkResult retireResult = m_frameContext.RetireCurrentCommandBuffer(submittingPreCommandBuffer);
         if (retireResult != VK_SUCCESS) {
             MGLOG_E_ONCE("FlushPendingCommands: RetireCurrentCommandBuffer returned %d; draining submission", retireResult);
+#if MOBILEGL_BUILD_DISAGGREGATED
+            if (WaitForSubmitsUpTo(m_submitCounter, UINT64_MAX)) {
+                // Every registered fence through this submission was waited.
+            } else if (vkQueueWaitIdle(m_graphicsQueue) == VK_SUCCESS) {
+#else
             if (vkWaitForFences(m_device, 1, &fence, VK_TRUE, UINT64_MAX) == VK_SUCCESS) {
                 OnSubmitsCompletedUpTo(m_submitCounter);
             } else if (vkQueueWaitIdle(m_graphicsQueue) == VK_SUCCESS) {
+#endif
                 m_bufferManager.NotifyDeviceIdle();
                 OnSubmitsCompletedUpTo(m_submitCounter);
             } else {
@@ -14028,6 +14097,12 @@ void main() {
                 return false;
             }
         }
+#if MOBILEGL_BUILD_DISAGGREGATED
+        if (!WaitForSubmitsUpTo(submitIndex, timeoutNs)) return false;
+        // A blocking wait can drain a present-less loop's frame transients.
+        TryDrainFrameTransients();
+        return true;
+#else
         for (const auto& record : m_inFlightSubmits) {
             if (record.submitIndex >= submitIndex) {
                 const VkResult result = vkWaitForFences(m_device, 1, &record.fence, VK_TRUE, timeoutNs);
@@ -14048,6 +14123,7 @@ void main() {
         // No in-flight record at or beyond the index: it was already observed
         // complete via a fence wait on a later submission.
         return true;
+#endif
     }
 
     void VulkanRenderer::OnFrameCommandRecordingBegan(VkCommandBuffer commandBuffer) {
@@ -14196,6 +14272,15 @@ void main() {
                 return;
             }
             m_presentSuspended = false;
+#if MOBILEGL_BUILD_DISAGGREGATED
+            // A slot's Present fence alone does not prove its older pooled
+            // flushes retired. Wait the full prefix before FrameContext may
+            // free its retired command buffers and reset the slot fence.
+            const Uint64 slotSubmit = m_frameContext.GetCurrent().lastSubmitIndex;
+            if (slotSubmit > m_completedSubmitCounter &&
+                !WaitForSubmitsUpTo(slotSubmit, UINT64_MAX))
+                MagmaWireFatal("deferred-acquire-submit-wait");
+#endif
             const VkResult acquireResult =
                 m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
             if (acquireResult == VK_SUBOPTIMAL_KHR) {
@@ -14326,6 +14411,16 @@ void main() {
         // 3) Advance frame slot.
         m_frameContext.AdvanceToNext();
 
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // FrameContext waits then resets the slot fence and frees retired
+        // command buffers. A later Present fence is not aggregate proof for
+        // earlier pooled flushes, so wait every registered submission in this
+        // slot's prefix BEFORE the reset/free happens.
+        const Uint64 slotSubmit = m_frameContext.GetCurrent().lastSubmitIndex;
+        if (slotSubmit > m_completedSubmitCounter &&
+            !WaitForSubmitsUpTo(slotSubmit, UINT64_MAX))
+            MagmaWireFatal("present-slot-submit-wait");
+#endif
         // 4) Wait/reset/acquire for next frame.
         result = m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
         if (result == VK_SUBOPTIMAL_KHR) {
@@ -14350,10 +14445,9 @@ void main() {
                 m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
         }
         VK_VERIFY(result, "Present, vkAcquireNextImageKHR");
-        // The acquired slot's fence has been waited: its last submission
+        // Pull keeps its historical slot-fence inference. In a disaggregated
-        // (and, in queue order, everything before it) is complete. The frame
+        // build the aggregate wait above already retired the registered prefix
-        // serials those submissions carried advance the buffer-manager floor
+        // before FrameContext reset the slot fence; this repeat is idempotent.
-        // inside OnSubmitsCompletedUpTo.
         OnSubmitsCompletedUpTo(m_frameContext.GetCurrent().lastSubmitIndex);
         CollectDeferredDepthMipmapCleanup(m_frameContext.GetCurrentFrameIndex());
 #if MOBILEGL_BUILD_DISAGGREGATED

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -147,7 +147,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 #if MOBILEGL_BUILD_DISAGGREGATED
         VkBufferManager& GetWireBufferManager() { return m_bufferManager; }
         Bool FlushWirePendingCommandsForTextureUpdate() {
-            return !HasPendingRecordedWork() || FlushPendingCommands();
+            // A new texture upload must not race graphics work that samples
+            // the old texels. No current recording does not imply no in-flight
+            // submission: finish the recorded work, then wait its whole prefix.
+            if (HasPendingRecordedWork() && !FlushPendingCommands()) return false;
+            return WaitForSubmitsUpTo(m_submitCounter, UINT64_MAX);
         }
         // P7 wave 2 package B3: the submission that will carry whatever is recorded NEXT.
         // RetireWireObjects (WireDraw.inc) already tags future objects with exactly this, and
@@ -685,6 +689,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // was waited or the device was idled); drops their records and
         // recycles pooled fences.
         void OnSubmitsCompletedUpTo(Uint64 submitIndex);
+#if MOBILEGL_BUILD_DISAGGREGATED
+        // A later fence alone cannot retire earlier submissions. Wait for the
+        // whole still-live prefix before advancing the completed submit floor.
+        Bool WaitForSubmitsUpTo(Uint64 submitIndex, Uint64 timeoutNs);
+#endif
         VkFence AcquirePooledSubmitFence();
         void DestroySubmitFencePool();
         Bool HasPendingRecordedWork() const;

MobileGL/MG_Remote/Server/StagedTextureStore.h

@@ -236,6 +236,16 @@ namespace MobileGL::MG_Remote::Server {
             if (!m_copies) return nullptr;
             const std::lock_guard<std::mutex> lock(m_mutex);
             LevelShadow& shadow = m_shadows[key].Levels[PackLevel(uploadTarget, level)];
+#if MOBILEGL_BUILD_DISAGGREGATED
+            // A run can be the first piece of a redefined level. Old coverage
+            // belongs to the previous coordinate system even when the new run
+            // happens to touch its old range; retaining it invents bytes that
+            // the current definition never supplied.
+            if (!shadow.Defined || shadow.Extent != extent) {
+                shadow.Bytes.clear();
+                shadow.Covered.clear();
+            }
+#endif
             shadow.Extent = extent;
             shadow.Defined = true;
             shadow.GpuDirty = false;

MobileGL/MG_Test/Wire/StagedTextureStoreTest.cpp

@@ -307,6 +307,27 @@ TEST(StagedTextureStoreTest, StageChunkRunsAssembleTheLevelImageAndAGapIsNotCove
     EXPECT_EQ(store.LevelByteSize(key, kTex2DTarget, 0), 0u);
 }
 
+#if MOBILEGL_BUILD_DISAGGREGATED
+TEST(StagedTextureStoreTest, AdoptRunExtentMoveDoesNotInheritOldCoverage) {
+    Server::StagedTextureStore store(/*copies=*/true);
+    const Uint64 key = Server::StagedTextureStore::KeyForHandle(TestHandle(15, 1));
+    Vector<Uint8> oldRun(16, 0x44);
+    Vector<Uint8> newRun(16, 0x99);
+    store.AdoptRun(key, kTex2DTarget, 0, IntVec3{4, 4, 1}, 0, oldRun.data(), oldRun.size());
+    ASSERT_TRUE(store.IsCovered(key, kTex2DTarget, 0));
+    store.AdoptRun(key, kTex2DTarget, 0, IntVec3{8, 4, 1}, 16, newRun.data(), newRun.size());
+    EXPECT_FALSE(store.IsCovered(key, kTex2DTarget, 0))
+        << "the old level's [0,16) run was mistaken for bytes of the new extent";
+    EXPECT_EQ(store.LevelCoveredRunCount(key, kTex2DTarget, 0), 1u);
+    Vector<Uint8> newLower(16, 0xAA);
+    store.AdoptRun(key, kTex2DTarget, 0, IntVec3{8, 4, 1}, 0, newLower.data(), newLower.size());
+    const Uint8* base = store.RequireLevelBytes(key, kTex2DTarget, 0, "unit_extent_move");
+    ASSERT_NE(base, nullptr);
+    EXPECT_EQ(base[0], 0xAAu);
+    EXPECT_EQ(base[16], 0x99u);
+}
+#endif
+
 // WHICH OF THE TWO STAGED-RUN SHAPES A RECORD CARRIES, read off the run's own LENGTH because the
 // wire layout is frozen and carries no field for it. The whole-level run (the shape this half has
 // always staged) is the level's complete shadow and begins at the level's first byte; a slab's run

docs/Disaggregated/notes/p7/magma-b4.md

@@ -0,0 +1,21 @@
+# P7 B4：Magma 提交等待与长帧清理（草稿，未落地）
+
+基线：`0e16ca27`，独立 worktree `~/w7/p7-b4`。本包仍在实现／审查，以下是当前源码状态，不是集成树结论。
+
+## 提交与 frame serial
+
+`WaitForSubmitsUpTo(index, timeout)` 收集 tracker 中所有 `submitIndex <= index` 的在飞 fence，以 `vkWaitForFences(..., VK_TRUE)` 等全部信号之后才推进完成计数。`WaitForSubmitIndex` 不再拿一条较晚 fence 代替整个前缀。`WaitForFrameSerial` 等待第一条匹配记录后重查地板；若同一 serial 仍被第二条提交持有，就从头重扫（不能在 `OnSubmitsCompletedUpTo` erase 后继续旧迭代器，也不再读 `record` 引用）。没有可用记录时才 queue idle，最终只按 `IsFrameSerialComplete` 答 true。
+
+`Present` 在 `FrameContext::WaitAndAcquireNextImage` 重置槽 fence / 释放退休命令缓冲之前，先等该槽 `lastSubmitIndex` 的完整前缀；suspended 后首次 acquire 同样做。readback 原本已有全 fence 等待，保持；mid-frame flush 的退化重用分支改走聚合等待。`HasPendingRecordedWork` 在 disagg 构建下也计 pre-pass 录制，`FlushPendingCommands` 的空批判断对应扩展。这样 `TryDrainFrameTransients` 的 wire `CollectWireObjects(all=true)` 不会在 pre-pass 尚未提交时把未来标签视作空闲。pull 分支的可执行语句保持原样。
+
+`FlushWirePendingCommandsForTextureUpdate` 在提交当前录制后等所有在飞 graphics 提交，才允许新的独立上传批修改旧 draw 可能采样的图像。这里可能有明显的 stall 代价，按规则只记录性能，不设门；需在长 trace 上实测。该修正还没有针对单 fence 错退的确定性负控，当前包不得以现有绿数直接落地。
+
+## 长帧回收与 staged bytes
+
+wire draw / dispatch 入口现在调用 `VkTextureManager::CollectGarbage`，使每 256 次的 `pruneWire` 能在一帧内执行。`StagedTextureStore::AdoptRun` 遇到 extent 变化时清旧 `Bytes` 与 `Covered`：单测先送旧 [0,16)，再送新 extent 的 [16,32)，不能把两种坐标系拼成已覆盖。禁用清理时该单测在旧范围误判覆盖处红，恢复后绿。
+
+## 已验证／待做
+
+完整 `MobileGL`、`MobileGLServer`、`MobileGLIntegrationTest` 与 `StagedTextureStoreTest` 构建通过；新 extent 单测绿，红一次按上述覆盖断言红；Magma split/spawn/tcp 为 101/80/71、full-split 539，全绿；OpenRA DirectVulkan inproc / spawn 都 ssim 1.000000、0 Fatal、0 `unsound-serial-complete`。原始日志 `~/w7/logs/p7b4/`、`~/w7/logs/p7-gate-b4a-retrace/`。
+
+待做：聚合等待的可证伪负控；G1 / 普查 / 棘轮 / parity / 完整八车道 / verify 门；集成者对 pipe 源码复审；与 M3 的 `VulkanRenderer.cpp` 冲突合并；真机回归。未完成前不推送本包。