LivoxGen1: Use syncCancelerForAsyncWork in producer pipeline

2026-05-29 14:10:45 -04:00
parent 5a9fe12057
commit d788810a05
7 changed files with 507 additions and 442 deletions
@@ -464,117 +464,137 @@ public:
 	void produceFrameReq1_doAssemble_posted(
 		std::shared_ptr<ProduceFrameReq> context)
 	{
-		sscl::SpinLock::Guard lock(pcloudProducer.shouldContinueLock);
-		if (!pcloudProducer.shouldContinue)
+		const bool shouldContinue = pcloudProducer.stimulusProducerCanceler
+			.execUncancelableSegmentOrAbort(
+				[this, context]()
+			{
+				pcloudProducer.ioUringAssemblyEngine.assembleFrameReq(
+					{context, std::bind(
+						&ProduceFrameReq::produceFrameReq2_assembleDone,
+						context.get(), context,
+						std::placeholders::_1, std::placeholders::_2)});
+			});
+
+		if (!shouldContinue)
 		{
 			callOriginalCallback();
 			return;
 		}
-
-		pcloudProducer.ioUringAssemblyEngine.assembleFrameReq(
-			{context, std::bind(
-				&ProduceFrameReq::produceFrameReq2_assembleDone,
-				context.get(), context,
-				std::placeholders::_1, std::placeholders::_2)});
 	}

 	void produceFrameReq2_assembleDone(
 		std::shared_ptr<ProduceFrameReq> context,
 		bool success, sscl::AsynchronousLoop loop)
 	{
-		sscl::SpinLock::Guard lock(pcloudProducer.shouldContinueLock);
-		if (!pcloudProducer.shouldContinue)
+		bool shouldContinue = pcloudProducer.stimulusProducerCanceler
+			.execUncancelableSegmentOrAbort(
+				[this, context, success, loop]()
 		{
-			callOriginalCallback();
-			return;
-		}
+			if (!success)
+			{
+				callOriginalCallback();

-		if (!success)
-		{
-			callOriginalCallback();
-
-			if (pcloudProducer.attachedStimulusBuffers.size() > 0) {
-				std::cerr << __func__ << ": Failed to assemble frame.\n";
+				if (pcloudProducer.attachedStimulusBuffers.size() > 0) {
+					std::cerr << __func__
+						<< ": Failed to assemble frame.\n";
+				}
+				return;
 			}
+
+			context->frameAssemblyResult = loop;
+
+			// Check if intensity buffer is attached and acquire frame if so
+			if (auto intensityBuff = pcloudProducer
+				.intensityStimulusBuffer.load(
+					std::memory_order_acquire))
+			{
+				size_t intensityRingbuffIndex = intensityBuff
+					->ringBuffer.getIndexToProduceInto();
+
+				StimulusFrame& intensityStimFrame = intensityBuff
+					->ringBuffer.getDataAtSlot(
+						intensityRingbuffIndex);
+
+				intensityStimFrame.lock.writeAcquire();
+				context->intensityStimFrame = std::make_optional(
+					std::ref(intensityStimFrame));
+			}
+			else {
+				context->intensityStimFrame = std::nullopt;
+			}
+
+			// Check if light ambience buffer is attached and acquire frame if so
+			std::optional<AmbienceProductionDesc>
+				lightAmbienceProductionDescDesc;
+			if (auto lightAmbienceBuff =
+				pcloudProducer.lightAmbienceStimulusBuffer.load(
+					std::memory_order_acquire))
+			{
+				size_t lightAmbienceRingbuffIndex = lightAmbienceBuff
+					->ringBuffer.getIndexToProduceInto();
+
+				StimulusFrame& lightAmbienceStimFrame =
+					lightAmbienceBuff->ringBuffer.getDataAtSlot(
+						lightAmbienceRingbuffIndex);
+
+				lightAmbienceStimFrame.lock.writeAcquire();
+				context->lightAmbienceStimFrame = std::make_optional(
+					std::ref(lightAmbienceStimFrame));
+				lightAmbienceProductionDescDesc =
+					AmbienceProductionDesc{
+						std::ref(lightAmbienceStimFrame),
+						lightAmbienceBuff
+							->passbandCountGtComparator};
+			}
+			else {
+				context->lightAmbienceStimFrame = std::nullopt;
+			}
+
+			// Check if dark ambience buffer is attached and acquire frame if so
+			std::optional<AmbienceProductionDesc>
+				darkAmbienceProductionDescDesc;
+			if (auto darkAmbienceBuff =
+				pcloudProducer.darkAmbienceStimulusBuffer.load(
+					std::memory_order_acquire))
+			{
+				size_t darkAmbienceRingbuffIndex = darkAmbienceBuff
+					->ringBuffer.getIndexToProduceInto();
+
+				StimulusFrame& darkAmbienceStimFrame =
+					darkAmbienceBuff->ringBuffer.getDataAtSlot(
+						darkAmbienceRingbuffIndex);
+
+				darkAmbienceStimFrame.lock.writeAcquire();
+				context->darkAmbienceStimFrame = std::make_optional(
+					std::ref(darkAmbienceStimFrame));
+				darkAmbienceProductionDescDesc =
+					AmbienceProductionDesc{
+						std::ref(darkAmbienceStimFrame),
+						darkAmbienceBuff
+							->passbandCountLtComparator};
+			}
+			else {
+				context->darkAmbienceStimFrame = std::nullopt;
+			}
+
+			pcloudProducer.openClCollatingAndMeshingEngine
+				.compactCollateAndMeshFrameReq(
+					context->frameAssemblyResult, stimulusFrame,
+					context->intensityStimFrame,
+					std::move(lightAmbienceProductionDescDesc),
+					std::move(darkAmbienceProductionDescDesc),
+					{context, std::bind(
+						&ProduceFrameReq
+							::produceFrameReq3_compactCollateDone,
+						context.get(), context,
+						std::placeholders::_1, std::placeholders::_2)});
+		});
+
+		if (!shouldContinue)
+		{
+			callOriginalCallback();
 			return;
 		}
-
-		context->frameAssemblyResult = loop;
-
-		// Check if intensity buffer is attached and acquire frame if so
-		if (auto intensityBuff = pcloudProducer.intensityStimulusBuffer.load(
-			std::memory_order_acquire))
-		{
-			size_t intensityRingbuffIndex = intensityBuff
-				->ringBuffer.getIndexToProduceInto();
-
-			StimulusFrame& intensityStimFrame = intensityBuff
-				->ringBuffer.getDataAtSlot(
-					intensityRingbuffIndex);
-
-			intensityStimFrame.lock.writeAcquire();
-			context->intensityStimFrame = std::make_optional(
-				std::ref(intensityStimFrame));
-		}
-		else {
-			context->intensityStimFrame = std::nullopt;
-		}
-
-		// Check if light ambience buffer is attached and acquire frame if so
-		std::optional<AmbienceProductionDesc> lightAmbienceProductionDescDesc;
-		if (auto lightAmbienceBuff =
-			pcloudProducer.lightAmbienceStimulusBuffer.load(
-				std::memory_order_acquire))
-		{
-			size_t lightAmbienceRingbuffIndex = lightAmbienceBuff
-				->ringBuffer.getIndexToProduceInto();
-
-			StimulusFrame& lightAmbienceStimFrame = lightAmbienceBuff
-				->ringBuffer.getDataAtSlot(lightAmbienceRingbuffIndex);
-
-			lightAmbienceStimFrame.lock.writeAcquire();
-			context->lightAmbienceStimFrame = std::make_optional(
-				std::ref(lightAmbienceStimFrame));
-			lightAmbienceProductionDescDesc = AmbienceProductionDesc{
-				std::ref(lightAmbienceStimFrame),
-				lightAmbienceBuff->passbandCountGtComparator};
-		}
-		else {
-			context->lightAmbienceStimFrame = std::nullopt;
-		}
-
-		// Check if dark ambience buffer is attached and acquire frame if so
-		std::optional<AmbienceProductionDesc> darkAmbienceProductionDescDesc;
-		if (auto darkAmbienceBuff =
-			pcloudProducer.darkAmbienceStimulusBuffer.load(
-				std::memory_order_acquire))
-		{
-			size_t darkAmbienceRingbuffIndex = darkAmbienceBuff
-				->ringBuffer.getIndexToProduceInto();
-
-			StimulusFrame& darkAmbienceStimFrame = darkAmbienceBuff
-				->ringBuffer.getDataAtSlot(darkAmbienceRingbuffIndex);
-
-			darkAmbienceStimFrame.lock.writeAcquire();
-			context->darkAmbienceStimFrame = std::make_optional(
-				std::ref(darkAmbienceStimFrame));
-			darkAmbienceProductionDescDesc = AmbienceProductionDesc{
-				std::ref(darkAmbienceStimFrame),
-				darkAmbienceBuff->passbandCountLtComparator};
-		}
-		else {
-			context->darkAmbienceStimFrame = std::nullopt;
-		}
-
-		pcloudProducer.openClCollatingAndMeshingEngine.compactCollateAndMeshFrameReq(
-			loop, stimulusFrame,
-			context->intensityStimFrame,
-			std::move(lightAmbienceProductionDescDesc),
-			std::move(darkAmbienceProductionDescDesc),
-			{context, std::bind(
-				&ProduceFrameReq::produceFrameReq3_compactCollateDone,
-				context.get(), context,
-				std::placeholders::_1, std::placeholders::_2)});
 	}

 	void produceFrameReq3_compactCollateDone(
@@ -603,7 +623,10 @@ public:
 		}
 #endif

-		// Release intensity frame if it was used
+		/**	EXPLANATION:
+		 * Release intensity/ambience frames if they were supplied/used,
+		 * regardless of whether or not a cancelation request occurred.
+		 */
 		if (context->intensityStimFrame.has_value()) {
 			context->intensityStimFrame->get().lock.writeRelease();
 		}
@@ -615,130 +638,169 @@ public:
 			context->darkAmbienceStimFrame->get().lock.writeRelease();
 		}

-		sscl::SpinLock::Guard lock(pcloudProducer.shouldContinueLock);
-		if (!pcloudProducer.shouldContinue)
+		if (!success)
+		{
+			callOriginalCallback();
+
+			std::cerr << __func__
+				<< ": Failed to compact and collate frame"
+				<< std::endl;
+
+			return;
+		}
+
+		/**	EXPLANATION:
+		 * Cancellation early exit for the success path. Analogous to the
+		 * (legacy, removed) pre-canceler check under shouldContinueLock before
+		 * lock.unlockPrematurely() in the former CPS version of this handler.
+		 *
+		 * Assumptions that make the unlocked tail (dump, debug logging, stage
+		 * durations, final callOriginalCallback()) safe without holding
+		 * stimulusProducerCanceler.s.lock:
+		 *
+		 * 1. This handler is only entered from
+		 *    CompactCollateAndMeshFrameReq4's execUncancelableSegmentOrAbort
+		 *    body, so openClCollatingAndMeshingEngine::finalize() cannot
+		 *    tear down OpenCL/engine state until this function returns.
+		 *
+		 * 2. PcloudStimulusProducer::stop() finalizes OpenCL before io_uring,
+		 *    so getCompactKernelDuration()/getCollateKernelDuration() and
+		 *    getAssemblyDuration() are not racing engine finalize here.
+		 *
+		 * 3. dump/logging reads only producer-owned state (pcloudFrameDumper,
+		 *    collationBuffer, device, context->frameAssemblyResult) that
+		 *    stop()/finalize() do not destroy.
+		 *
+		 * 4. A stop() that lands after this read but before the tail is
+		 *    intentional stale work (same as unlockPrematurely()), not UAF.
+		 *
+		 * Rework to run portions inside stimulusProducerCanceler
+		 * execUncancelableSegmentOrAbort if any of the above ceases to hold:
+		 * e.g. this callback is invoked outside the OpenCL uncancelable
+		 * segment, the tail begins touching engine or buffer state that
+		 * finalize() resets, or stop() ordering changes so teardown can
+		 * interleave with the unlocked tail.
+		 */
+		const bool shouldContinue =
+			!pcloudProducer.stimulusProducerCanceler.isCancellationRequested();
+
+		if (!shouldContinue)
 		{
 			callOriginalCallback();
 			return;
 		}

-		if (!success) {
-			std::cerr << __func__ << ": Failed to compact and collate frame" << std::endl;
-		} else
+		if (pcloudProducer.pcloudFrameDumper.isEnabled())
 		{
-			lock.unlockPrematurely();
-			if (pcloudProducer.pcloudFrameDumper.isEnabled())
+			try
 			{
-				try
-				{
-					pcloudProducer.pcloudFrameDumper.dumpProducedFrame(
-						*pcloudProducer.device,
-						pcloudProducer.collationBuffer,
-						context->frameAssemblyResult);
-				}
-				catch (const std::exception& e)
-				{
-					std::cerr << __func__ << ": Failed to dump pcloud frame: "
-						<< e.what() << std::endl;
-				}
+				pcloudProducer.pcloudFrameDumper.dumpProducedFrame(
+					*pcloudProducer.device,
+					pcloudProducer.collationBuffer,
+					context->frameAssemblyResult);
 			}
+			catch (const std::exception& e)
+			{
+				std::cerr << __func__ << ": Failed to dump pcloud frame: "
+					<< e.what() << std::endl;
+			}
+		}

 #if SMO_DEBUG_PCLOUD_AMBIENCE_INTRIN
-			if (logLightAmbience)
+		if (logLightAmbience)
+		{
+			auto lightBuff =
+				pcloudProducer.lightAmbienceStimulusBuffer.load(
+					std::memory_order_acquire);
+			if (lightBuff)
 			{
-				auto lightBuff =
-					pcloudProducer.lightAmbienceStimulusBuffer.load(
-						std::memory_order_acquire);
-				if (lightBuff)
+				std::cerr << __func__ << ": pcloudLightAmbience "
+					<< "passbandCount=" << logLightPassbandCount
+					<< " (per-slot avg intensity "
+					<< ambienceComparatorOpChar(
+						lightBuff->passbandCountGtComparator.op)
+					<< " " << lightBuff->passbandCountGtComparator.value
+					<< ")";
+				if (lightBuff->negtrinInterestConfig.has_value())
 				{
-					std::cerr << __func__ << ": pcloudLightAmbience "
-						<< "passbandCount=" << logLightPassbandCount
-						<< " (per-slot avg intensity "
-						<< ambienceComparatorOpChar(
-							lightBuff->passbandCountGtComparator.op)
-						<< " " << lightBuff->passbandCountGtComparator.value
-						<< ")";
-					if (lightBuff->negtrinInterestConfig.has_value())
+					const auto& nc = *lightBuff->negtrinInterestConfig;
+					std::cerr << " negtrinInterestThr=" << nc.threshold;
+					if (nc.percentage != 0U)
 					{
-						const auto& nc = *lightBuff->negtrinInterestConfig;
-						std::cerr << " negtrinInterestThr=" << nc.threshold;
-						if (nc.percentage != 0U)
-						{
-							std::cerr << " (from " << nc.percentage << "%)";
-						}
-						std::cerr << " meetsNegtrinInterest="
-							<< (lightBuff->shouldTriggerNegtrinEvent(
-								logLightPassbandCount)
-								? "yes" : "no");
+						std::cerr << " (from " << nc.percentage << "%)";
 					}
-					else
-					{
-						std::cerr << " negtrinInterest(n/a)";
-					}
-					std::cerr << std::endl;
+					std::cerr << " meetsNegtrinInterest="
+						<< (lightBuff->shouldTriggerNegtrinEvent(
+							logLightPassbandCount)
+							? "yes" : "no");
 				}
+				else
+				{
+					std::cerr << " negtrinInterest(n/a)";
+				}
+				std::cerr << std::endl;
 			}
-			if (logDarkAmbience)
+		}
+		if (logDarkAmbience)
+		{
+			auto darkBuff =
+				pcloudProducer.darkAmbienceStimulusBuffer.load(
+					std::memory_order_acquire);
+			if (darkBuff)
 			{
-				auto darkBuff =
-					pcloudProducer.darkAmbienceStimulusBuffer.load(
-						std::memory_order_acquire);
-				if (darkBuff)
+				std::cerr << __func__ << ": pcloudDarkAmbience "
+					<< "passbandCount=" << logDarkPassbandCount
+					<< " (per-slot avg intensity "
+					<< ambienceComparatorOpChar(
+						darkBuff->passbandCountLtComparator.op)
+					<< " " << darkBuff->passbandCountLtComparator.value
+					<< ")";
+				if (darkBuff->postrinInterestConfig.has_value())
 				{
-					std::cerr << __func__ << ": pcloudDarkAmbience "
-						<< "passbandCount=" << logDarkPassbandCount
-						<< " (per-slot avg intensity "
-						<< ambienceComparatorOpChar(
-							darkBuff->passbandCountLtComparator.op)
-						<< " " << darkBuff->passbandCountLtComparator.value
-						<< ")";
-					if (darkBuff->postrinInterestConfig.has_value())
+					const auto& pc = *darkBuff->postrinInterestConfig;
+					std::cerr << " postrinInterestThr=" << pc.threshold;
+					if (pc.percentage != 0U)
 					{
-						const auto& pc = *darkBuff->postrinInterestConfig;
-						std::cerr << " postrinInterestThr=" << pc.threshold;
-						if (pc.percentage != 0U)
-						{
-							std::cerr << " (from " << pc.percentage << "%)";
-						}
-						std::cerr << " meetsPostrinInterest="
-							<< (darkBuff->shouldTriggerPostrinEvent(
-								logDarkPassbandCount)
-								? "yes" : "no");
+						std::cerr << " (from " << pc.percentage << "%)";
 					}
-					else
-					{
-						std::cerr << " postrinInterest(n/a)";
-					}
-					std::cerr << std::endl;
+					std::cerr << " meetsPostrinInterest="
+						<< (darkBuff->shouldTriggerPostrinEvent(
+							logDarkPassbandCount)
+							? "yes" : "no");
 				}
+				else
+				{
+					std::cerr << " postrinInterest(n/a)";
+				}
+				std::cerr << std::endl;
 			}
+		}
 #endif

 #if SMO_PRINT_PCLOUD_STAGE_DURATIONS
-			const auto logNow = std::chrono::system_clock::now();
-			const std::time_t logTime =
-				std::chrono::system_clock::to_time_t(logNow);
-			const auto logSubsecMs =
-				std::chrono::duration_cast<std::chrono::milliseconds>(
-					logNow.time_since_epoch()) % 1000;
-			auto assemblyDuration =
-				pcloudProducer.ioUringAssemblyEngine.getAssemblyDuration();
-			auto compactDuration =
-				pcloudProducer.openClCollatingAndMeshingEngine
-					.getCompactKernelDuration();
-			auto collateDuration =
-				pcloudProducer.openClCollatingAndMeshingEngine
-					.getCollateKernelDuration();
-			std::cout << std::put_time(std::localtime(&logTime), "%T")
-				<< '.' << std::setfill('0') << std::setw(3)
-				<< logSubsecMs.count() << ' '
-				<< __func__ << ": stage durations: assembly="
-				<< assemblyDuration.count()
-				<< "ms, compactKernel=" << compactDuration.count()
-				<< "ms, collateKernel=" << collateDuration.count()
-				<< "ms" << std::endl;
+		const auto logNow = std::chrono::system_clock::now();
+		const std::time_t logTime =
+			std::chrono::system_clock::to_time_t(logNow);
+		const auto logSubsecMs =
+			std::chrono::duration_cast<std::chrono::milliseconds>(
+				logNow.time_since_epoch()) % 1000;
+		auto assemblyDuration =
+			pcloudProducer.ioUringAssemblyEngine.getAssemblyDuration();
+		auto compactDuration =
+			pcloudProducer.openClCollatingAndMeshingEngine
+				.getCompactKernelDuration();
+		auto collateDuration =
+			pcloudProducer.openClCollatingAndMeshingEngine
+				.getCollateKernelDuration();
+		std::cout << std::put_time(std::localtime(&logTime), "%T")
+			<< '.' << std::setfill('0') << std::setw(3)
+			<< logSubsecMs.count() << ' '
+			<< __func__ << ": stage durations: assembly="
+			<< assemblyDuration.count()
+			<< "ms, compactKernel=" << compactDuration.count()
+			<< "ms, collateKernel=" << collateDuration.count()
+			<< "ms" << std::endl;
 #endif
-		}

 		callOriginalCallback();
 	}
@@ -748,10 +810,9 @@ void PcloudStimulusProducer::produceFrameReq(
 	sscl::cps::Callback<produceFrameReqCbFn> callback)
 {
 	/**	EXPLANATION:
-	 * We shouldn't acquire the StimulusProducer::shouldContinueLock here because
-	 * this function is called from
-	 * StimulusProducer::stimFrameProductionTimesliceInd(), which is already
-	 * holding the lock.
+	 * We don't do any additional canceler-lock acquisition here because
+	 * callback segment methods already use stimulusProducerCanceler
+	 * checkpoints before running uncancelable work.
 	 */
 	auto caller = smoHooksPtr->ComponentThread_getSelf();
 	auto request = std::make_shared<ProduceFrameReq>(