From 84555746156296257ebbff6aab85a19d49668f75 Mon Sep 17 00:00:00 2001
From: psucien <bad_cast@protonmail.com>
Date: Fri, 17 May 2024 23:31:19 +0200
Subject: [PATCH] gnmdriver: submission lock moved out from gpu

---
 src/core/libraries/gnmdriver/gnmdriver.cpp | 27 +++++++++++++++-------
 src/video_core/amdgpu/liverpool.h          | 16 ++-----------
 2 files changed, 21 insertions(+), 22 deletions(-)

diff --git a/src/core/libraries/gnmdriver/gnmdriver.cpp b/src/core/libraries/gnmdriver/gnmdriver.cpp
index 5cb7f5a9..b79ef820 100644
--- a/src/core/libraries/gnmdriver/gnmdriver.cpp
+++ b/src/core/libraries/gnmdriver/gnmdriver.cpp
@@ -23,6 +23,9 @@ static std::unique_ptr<AmdGpu::Liverpool> liverpool;
 // support is not important and can be ignored for a while.
 static constexpr bool g_fair_hw_init = false;
 
+// In case if `submitDone` is issued we need to block submissions until GPU idle
+static u32 submission_lock{};
+
 // Write a special ending NOP packet with N DWs data block
 template <u32 data_block_size>
 static inline u32* WriteTrailingNop(u32* cmdbuf) {
@@ -50,18 +53,20 @@ s32 PS4_SYSV_ABI sceGnmAddEqEvent(SceKernelEqueue eq, u64 id, void* udata) {
     eq->addEvent(kernel_event);
 
     Platform::IrqC::Instance()->Register(
-        Platform::InterruptId::GfxEop, [=](Platform::InterruptId irq) {
+        Platform::InterruptId::GfxEop,
+        [=](Platform::InterruptId irq) {
             ASSERT_MSG(irq == Platform::InterruptId::GfxEop,
                        "An unexpected IRQ occured"); // We need to conver IRQ# to event id and do
                                                      // proper filtering in trigger function
             eq->triggerEvent(SceKernelEvent::Type::GfxEop, EVFILT_GRAPHICS_CORE, nullptr);
-        });
+        },
+        eq);
     return ORBIS_OK;
 }
 
 int PS4_SYSV_ABI sceGnmAreSubmitsAllowed() {
-    LOG_ERROR(Lib_GnmDriver, "(STUBBED) called");
-    return ORBIS_OK;
+    LOG_TRACE(Lib_GnmDriver, "called");
+    return submission_lock == 0;
 }
 
 int PS4_SYSV_ABI sceGnmBeginWorkload() {
@@ -165,7 +170,7 @@ s32 PS4_SYSV_ABI sceGnmDeleteEqEvent(SceKernelEqueue eq, u64 id) {
 
     eq->removeEvent(id);
 
-    Platform::IrqC::Instance()->Unregister(Platform::InterruptId::GfxEop);
+    Platform::IrqC::Instance()->Unregister(Platform::InterruptId::GfxEop, eq);
     return ORBIS_OK;
 }
 
@@ -1411,6 +1416,14 @@ s32 PS4_SYSV_ABI sceGnmSubmitCommandBuffers(u32 count, const u32* dcb_gpu_addrs[
         }
     }
 
+    if (submission_lock != 0) {
+        liverpool->WaitGpuIdle();
+
+        // Suspend logic goes here
+
+        submission_lock = 0;
+    }
+
     for (auto cbpair = 0u; cbpair < count; ++cbpair) {
         const auto* ccb = ccb_gpu_addrs ? ccb_gpu_addrs[cbpair] : nullptr;
         const auto ccb_size = ccb_sizes_in_bytes ? ccb_sizes_in_bytes[cbpair] : 0;
@@ -1428,9 +1441,7 @@ int PS4_SYSV_ABI sceGnmSubmitCommandBuffersForWorkload() {
 
 int PS4_SYSV_ABI sceGnmSubmitDone() {
     LOG_INFO(Lib_GnmDriver, "called");
-
-    liverpool->SubmitDone();
-
+    submission_lock = true;
     return ORBIS_OK;
 }
 
diff --git a/src/video_core/amdgpu/liverpool.h b/src/video_core/amdgpu/liverpool.h
index c8e8eb0d..b40c9ba5 100644
--- a/src/video_core/amdgpu/liverpool.h
+++ b/src/video_core/amdgpu/liverpool.h
@@ -620,14 +620,6 @@ public:
     ~Liverpool();
 
     void SubmitGfx(std::span<const u32> dcb, std::span<const u32> ccb) {
-        if (submission_lock) {
-            WaitGpuIdle();
-
-            // Suspend logic goes here
-
-            submission_lock = false;
-        }
-
         {
             std::scoped_lock lock{m_ring_access};
             gfx_ring.emplace(dcb);
@@ -636,22 +628,18 @@ public:
         }
         cv_submit.notify_one();
     }
-    void SubmitDone() {
-        submission_lock = true;
-    }
+
+    void WaitGpuIdle();
 
 private:
     void ProcessCmdList(const u32* cmdbuf, u32 size_in_bytes);
     void Process(std::stop_token stoken);
-    void WaitGpuIdle();
 
     std::jthread process_thread{};
     std::queue<std::span<const u32>> gfx_ring{};
     std::condition_variable_any cv_submit{};
     std::condition_variable cv_complete{};
     std::mutex m_ring_access{};
-
-    bool submission_lock{};
 };
 
 static_assert(GFX6_3D_REG_INDEX(ps_program) == 0x2C08);