msd_intel_gen.txt

sysdrv_bind()
    首先拿到i915_gpu_core_protocol_ops

    device->magma_driver = MagmaDriver::Create();
        msd_driver_create();
            MsdIntelDriver::Create().release();
                MsdIntelDriver()

    magma_start(device.get());
        device->magma_driver->CreateDevice(&device->gpu_core_protocol);
            msd_driver_create_device(msd_drv_.get(), device);
                MsdIntelDevice::Create(device_handle, start_device_thread);
                    device->Init(device_handle)
                        BaseInit(device_handle)
                            MsdIntelPciDevice::CreateShim(device_handle);
                                MsdIntelPciDeviceShim(zx_intel_gpu_core_protocol_t* intel_gpu_core)
                            
                            读取pci config里的device id, revision
                            映射pci设备的mmio

                            magma::PlatformBusMapper::Create(platform_device_->GetBusTransactionInitiator());
                                这个好像是映射内存用的？
                            platform_device_是 MsdIntelPciDeviceShim

                            QuerySliceInfo(&subslice_total_, &eu_total_);
                            InterruptManager::CreateShim(this);
                            PerProcessGtt::InitPrivatePat(register_io_.get());
                            Gtt::CreateShim(this)
                            Sequencer(kFirstSequenceNumber)

                            RenderEngineCommandStreamer::Create(this);
                                Scheduler::CreateFifoScheduler();
                            render_engine_cs_->InitContext(global_context_.get())
                                GetContextSize()
                                  // page size * 20
                                MsdIntelBuffer::Create(context_size, "context-buffer")
                                    magma::PlatformBuffer::Create(size, name);
                                    创建context buffer
                                        zx::vmo::create(size, 0, &vmo);
                                Ringbuffer(MsdIntelBuffer::Create(32 * PAGE_SIZE, "ring-buffer")
                                    创建ring buffer

                                InitContextBuffer(context_buffer.get(), ringbuffer.get(), context->exec_address_space().get()))    
                                    platform_buf->MapCpu(&addr)
                                        把vmo映射好，
                                        context buffer的布局在vol6 command stream programming, page 17
                                        第一个4k页是per-process hw status page.
                                        mmio_ = 8192的意思是：用0x02000h作为所有寄存器的base. 这样在指定地址时只需要指定低12位。

                                        register state context存放了一些mmio register的偏移量。

                                context->SetEngineState(id(), std::move(context_buffer), std::move(ringbuffer));
                                    做PerEngineState放入global_context

                            global_context_->Map(gtt_, render_engine_cs_->id())
                                地址空间类型ADDRESS_SPACE_GGTT
                                把context buffer, ring buffer映射给GGTT

                                MsdIntelContext::Map(address_space, id)
                                GetGpuAddress(id, &gpu_addr)
                                get_context_buffer(id)->platform_buffer()->MapCpu(&cpu_addr)

                                hardware status page实际上就是context buffer映射的ggtt offset

                        RenderEngineInit(true)
                            render_engine_cs_->InitHardware();
                                把hardware status page的gpu地址写到0x2080  Hardware Status Page Address Register
                                把sequence number写道hardware status page的offset 0x20处。这个在公开文档里写的是
                                Ring Head Pointer Storage。代码里注释是superNDA

                                execlist enable
                                设置各种寄存器


                            render_engine_cs_->CreateRenderInitBatch(device_id_);

                            render_engine_cs_->RenderInit(global_context_, std::move(init_batch), gtt_))
                                MsdIntelBuffer::Create(init_batch->size(), "render-init-batch"));
                                    新建一个buffer
                                init_batch->Init(std::move(buffer), address_space);
                                    用这个buffer来创建batch context
                                    这里要把新的地址写到相应的位置reloc

                                ExecBatch(std::move(mapped_batch));
                                    MoveBatchToInflight(std::move(mapped_batch)
                                        StartBatchBuffer(context.get(), gpu_addr, context->exec_address_space()->type())

                                    SubmitContext(context.get(), inflight_command_sequences_.back().ringbuffer_offset());
                                        UpdateContext(context, tail)
                                            tail是在ring buffer内部的tail offset
                                            gpuaddr是ring buffer在gtt里的偏移量

                                        SubmitExeclists(context);

                    device->StartDeviceThread();
                        std::thread([this] { this->DeviceThreadLoop(); });
                              线程循环处理请求队列