Add more fine-grained buffer recreation locking

app/src/main/cpp/skyline/gpu/buffer.cpp

@@ -697,6 +697,11 @@ namespace skyline::gpu {
         return delegate->GetBuffer();
     }
 
+    BufferBinding BufferView::GetBinding(GPU &gpu) const {
+        std::scoped_lock lock{gpu.buffer.recreationMutex};
+        return {delegate->GetBuffer()->GetBacking(), offset + delegate->GetOffset(), size};
+    }
+
     vk::DeviceSize BufferView::GetOffset() const {
         return offset + delegate->GetOffset();
     }

app/src/main/cpp/skyline/gpu/buffer.h

@@ -492,6 +492,13 @@ namespace skyline::gpu {
          */
         vk::DeviceSize GetOffset() const;
 
+        /**
+         * @return A binding describing the underlying buffer state at a given moment in time
+         * @note The view **must** be locked prior to calling this
+         * @note This is the **ONLY** function in BufferView that can be called from non-GPFIFO threads
+         */
+        BufferBinding GetBinding(GPU &gpu) const;
+
         /**
          * @note The buffer manager **MUST** be locked prior to calling this
          */

app/src/main/cpp/skyline/gpu/buffer_manager.h

@@ -69,7 +69,7 @@ namespace skyline::gpu {
         static bool BufferLessThan(const std::shared_ptr<Buffer> &it, u8 *pointer);
 
       public:
-        std::mutex recreationMutex;
+        SpinLock recreationMutex;
 
         BufferManager(GPU &gpu);
 

app/src/main/cpp/skyline/gpu/interconnect/command_executor.cpp

@@ -94,7 +94,6 @@ namespace skyline::gpu::interconnect {
         vk::RenderPass lRenderPass;
         u32 subpassIndex;
 
-        std::scoped_lock bufferLock{gpu.buffer.recreationMutex};
         using namespace node;
         for (NodeVariant &node : slot->nodes) {
             #define NODE(name) [&](name& node) { node(slot->commandBuffer, slot->cycle, gpu); }

app/src/main/cpp/skyline/gpu/interconnect/common/state_updater.h

@@ -66,9 +66,10 @@ namespace skyline::gpu::interconnect {
     struct SetVertexBuffersDynamicCmdImpl {
         void Record(GPU &gpu, vk::raii::CommandBuffer &commandBuffer) {
             for (u32 i{base.firstBinding}; i < base.firstBinding + base.bindingCount; i++) {
-                base.buffers[i] = views[i].GetBuffer()->GetBacking();
+                auto binding{views[i].GetBinding(gpu)};
-                base.offsets[i] = views[i].GetOffset();
+                base.buffers[i] = binding.buffer;
-                base.sizes[i] = views[i].size;
+                base.offsets[i] = binding.offset;
+                base.sizes[i] = binding.size;
             }
 
             base.Record(gpu, commandBuffer);
@@ -92,8 +93,9 @@ namespace skyline::gpu::interconnect {
 
     struct SetIndexBufferDynamicCmdImpl {
         void Record(GPU &gpu, vk::raii::CommandBuffer &commandBuffer) {
-            base.buffer = view.GetBuffer()->GetBacking();
+            auto binding{view.GetBinding(gpu)};
-            base.offset = view.GetOffset();
+            base.buffer = binding.buffer;
+            base.offset = binding.offset;
             base.Record(gpu, commandBuffer);
         }
 
@@ -116,9 +118,10 @@ namespace skyline::gpu::interconnect {
 
     struct SetTransformFeedbackBufferDynamicCmdImpl {
         void Record(GPU &gpu, vk::raii::CommandBuffer &commandBuffer) {
-            base.buffer = view.GetBuffer()->GetBacking();
+            auto binding{view.GetBinding(gpu)};
-            base.offset = view.GetOffset();
+            base.buffer = binding.buffer;
-            base.size = view.size;
+            base.offset = binding.offset;
+            base.size = binding.size;
             base.Record(gpu, commandBuffer);
         }
 
@@ -206,19 +209,20 @@ namespace skyline::gpu::interconnect {
             // Resolve descriptor infos from dynamic bindings
             for (size_t i{}; i < updateInfo->bufferDescDynamicBindings.size(); i++) {
                 auto &dynamicBinding{updateInfo->bufferDescDynamicBindings[i]};
-                if (auto view{std::get_if<BufferView>(&dynamicBinding)}) {
+                BufferBinding binding{[&dynamicBinding, &gpu]() {
-                    updateInfo->bufferDescs[i] = vk::DescriptorBufferInfo{
+                    if (auto view{std::get_if<BufferView>(&dynamicBinding)})
-                        .buffer = view->GetBuffer()->GetBacking(),
+                        return view->GetBinding(gpu);
-                        .offset = view->GetOffset(),
+                    else if (auto binding{std::get_if<BufferBinding>(&dynamicBinding)})
-                        .range = view->size
+                        return *binding;
-                    };
+                    else
-                } else if (auto binding{std::get_if<BufferBinding>(&dynamicBinding)}) {
+                        return BufferBinding{};
-                    updateInfo->bufferDescs[i] = vk::DescriptorBufferInfo{
+                }()};
-                        .buffer = binding->buffer,
+
-                        .offset = binding->offset,
+                updateInfo->bufferDescs[i] = vk::DescriptorBufferInfo{
-                        .range = binding->size
+                    .buffer = binding.buffer,
-                    };
+                    .offset = binding.offset,
-                }
+                    .range = binding.size
+                };
             }
 
             if constexpr (PushDescriptor) {

app/src/main/cpp/skyline/gpu/interconnect/inline2memory.cpp

@@ -28,13 +28,14 @@ namespace skyline::gpu::interconnect {
             dstBuf.GetBuffer()->BlockAllCpuBackingWrites();
 
             auto srcGpuAllocation{gpu.megaBufferAllocator.Push(executor.cycle, src)};
-            executor.AddOutsideRpCommand([srcGpuAllocation, dstBuf, src](vk::raii::CommandBuffer &commandBuffer, const std::shared_ptr<FenceCycle> &, GPU &) {
+            executor.AddOutsideRpCommand([srcGpuAllocation, dstBuf, src](vk::raii::CommandBuffer &commandBuffer, const std::shared_ptr<FenceCycle> &, GPU &pGpu) {
+                auto dstBufBinding{dstBuf.GetBinding(pGpu)};
                 vk::BufferCopy copyRegion{
                     .size = src.size_bytes(),
                     .srcOffset = srcGpuAllocation.offset,
-                    .dstOffset = dstBuf.GetOffset()
+                    .dstOffset = dstBufBinding.offset,
                 };
-                commandBuffer.copyBuffer(srcGpuAllocation.buffer, dstBuf.GetBuffer()->GetBacking(), copyRegion);
+                commandBuffer.copyBuffer(srcGpuAllocation.buffer, dstBufBinding.buffer, copyRegion);
                 commandBuffer.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer, vk::PipelineStageFlagBits::eAllCommands, {}, vk::MemoryBarrier{
                     .srcAccessMask = vk::AccessFlagBits::eTransferWrite,
                     .dstAccessMask = vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite

app/src/main/cpp/skyline/gpu/interconnect/maxwell_3d/constant_buffers.cpp

@@ -62,13 +62,14 @@ namespace skyline::gpu::interconnect::maxwell3d {
                 callbackData.view.GetBuffer()->BlockAllCpuBackingWrites();
 
                 auto srcGpuAllocation{callbackData.ctx.gpu.megaBufferAllocator.Push(callbackData.ctx.executor.cycle, callbackData.srcCpuBuf)};
-                callbackData.ctx.executor.AddOutsideRpCommand([=, srcCpuBuf = callbackData.srcCpuBuf, view = callbackData.view, offset = callbackData.offset](vk::raii::CommandBuffer &commandBuffer, const std::shared_ptr<FenceCycle> &, GPU &) {
+                callbackData.ctx.executor.AddOutsideRpCommand([=, srcCpuBuf = callbackData.srcCpuBuf, view = callbackData.view, offset = callbackData.offset](vk::raii::CommandBuffer &commandBuffer, const std::shared_ptr<FenceCycle> &, GPU &gpu) {
+                    auto binding{view.GetBinding(gpu)};
                     vk::BufferCopy copyRegion{
                         .size = srcCpuBuf.size_bytes(),
                         .srcOffset = srcGpuAllocation.offset,
-                        .dstOffset = view.GetOffset() + offset
+                        .dstOffset = binding.offset + offset
                     };
-                    commandBuffer.copyBuffer(srcGpuAllocation.buffer, view.GetBuffer()->GetBacking(), copyRegion);
+                    commandBuffer.copyBuffer(srcGpuAllocation.buffer, binding.buffer, copyRegion);
                     commandBuffer.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer, vk::PipelineStageFlagBits::eAllCommands, {}, vk::MemoryBarrier{
                         .srcAccessMask = vk::AccessFlagBits::eTransferWrite,
                         .dstAccessMask = vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite

app/src/main/cpp/skyline/gpu/interconnect/maxwell_dma.cpp

@@ -40,17 +40,19 @@ namespace skyline::gpu::interconnect {
             srcBuf.GetBuffer()->BlockAllCpuBackingWrites();
             dstBuf.GetBuffer()->BlockAllCpuBackingWrites();
 
-            executor.AddOutsideRpCommand([srcBuf, dstBuf](vk::raii::CommandBuffer &commandBuffer, const std::shared_ptr<FenceCycle> &, GPU &) {
+            executor.AddOutsideRpCommand([srcBuf, dstBuf](vk::raii::CommandBuffer &commandBuffer, const std::shared_ptr<FenceCycle> &, GPU &gpu) {
                 commandBuffer.pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands, vk::PipelineStageFlagBits::eTransfer, {}, vk::MemoryBarrier{
                     .srcAccessMask = vk::AccessFlagBits::eMemoryRead,
                     .dstAccessMask = vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite
                 }, {}, {});
+                auto srcBufBinding{srcBuf.GetBinding(gpu)};
+                auto dstBufBinding{dstBuf.GetBinding(gpu)};
                 vk::BufferCopy copyRegion{
                     .size = srcBuf.size,
-                    .srcOffset = srcBuf.GetOffset(),
+                    .srcOffset = srcBufBinding.offset,
-                    .dstOffset = dstBuf.GetOffset()
+                    .dstOffset = dstBufBinding.offset
                 };
-                commandBuffer.copyBuffer(srcBuf.GetBuffer()->GetBacking(), dstBuf.GetBuffer()->GetBacking(), copyRegion);
+                commandBuffer.copyBuffer(srcBufBinding.buffer, dstBufBinding.buffer, copyRegion);
                 commandBuffer.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer, vk::PipelineStageFlagBits::eAllCommands, {}, vk::MemoryBarrier{
                     .srcAccessMask = vk::AccessFlagBits::eTransferWrite,
                     .dstAccessMask = vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite,