9 files changed, 329 insertions, 27 deletions

diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
index ca54eb6c6..0f713ead1 100644
--- a/src/core/CMakeLists.txt
+++ b/src/core/CMakeLists.txt
@@ -611,6 +611,8 @@ add_library(core STATIC
     hle/service/ns/pdm_qry.h
     hle/service/nvdrv/core/container.cpp
     hle/service/nvdrv/core/container.h
+    hle/service/nvdrv/core/heap_mapper.cpp
+    hle/service/nvdrv/core/heap_mapper.h
     hle/service/nvdrv/core/nvmap.cpp
     hle/service/nvdrv/core/nvmap.h
     hle/service/nvdrv/core/syncpoint_manager.cpp
diff --git a/src/core/device_memory_manager.inc b/src/core/device_memory_manager.inc
index b3a5f3d8b..138eb5017 100644
--- a/src/core/device_memory_manager.inc
+++ b/src/core/device_memory_manager.inc
@@ -20,10 +20,10 @@ namespace Core {
 
 namespace {
 
-class PhysicalAddressContainer {
+class MultiAddressContainer {
 public:
-    PhysicalAddressContainer() = default;
-    ~PhysicalAddressContainer() = default;
+    MultiAddressContainer() = default;
+    ~MultiAddressContainer() = default;
 
     void GatherValues(u32 start_entry, Common::ScratchBuffer<u32>& buffer) {
         buffer.resize(8);
@@ -145,7 +145,7 @@ struct DeviceMemoryManagerAllocator {
     std::conditional_t<supports_pinning, Common::FlatAllocator<DAddr, 0, pin_bits>, EmptyAllocator>
         pin_allocator;
     Common::FlatAllocator<DAddr, 0, device_virtual_bits> main_allocator;
-    PhysicalAddressContainer multi_dev_address;
+    MultiAddressContainer multi_dev_address;
 
     /// Returns true when vaddr -> vaddr+size is fully contained in the buffer
     template <bool pin_area>
diff --git a/src/core/hle/service/nvdrv/core/container.cpp b/src/core/hle/service/nvdrv/core/container.cpp
index e12ce05c1..ba7eb9e24 100644
--- a/src/core/hle/service/nvdrv/core/container.cpp
+++ b/src/core/hle/service/nvdrv/core/container.cpp
@@ -8,6 +8,7 @@
 
 #include "core/hle/kernel/k_process.h"
 #include "core/hle/service/nvdrv/core/container.h"
+#include "core/hle/service/nvdrv/core/heap_mapper.h"
 #include "core/hle/service/nvdrv/core/nvmap.h"
 #include "core/hle/service/nvdrv/core/syncpoint_manager.h"
 #include "core/memory.h"
@@ -36,6 +37,14 @@ Container::~Container() = default;
 
 size_t Container::OpenSession(Kernel::KProcess* process) {
     std::scoped_lock lk(impl->session_guard);
+    for (auto& session : impl->sessions) {
+        if (!session.is_active) {
+            continue;
+        }
+        if (session.process == process) {
+            return session.id;
+        }
+    }
     size_t new_id{};
     auto* memory_interface = &process->GetMemory();
     auto& smmu = impl->host1x.MemoryManager();
@@ -48,16 +57,65 @@ size_t Container::OpenSession(Kernel::KProcess* process) {
         impl->sessions.emplace_back(new_id, process, smmu_id);
         new_id = impl->new_ids++;
     }
-    LOG_CRITICAL(Debug, "Created Session {}", new_id);
+    auto& session = impl->sessions[new_id];
+    session.is_active = true;
+    // Optimization
+    if (process->IsApplication()) {
+        auto& page_table = process->GetPageTable().GetBasePageTable();
+        auto heap_start = page_table.GetHeapRegionStart();
+
+        Kernel::KProcessAddress cur_addr = heap_start;
+        size_t region_size = 0;
+        VAddr region_start = 0;
+        while (true) {
+            Kernel::KMemoryInfo mem_info{};
+            Kernel::Svc::PageInfo page_info{};
+            R_ASSERT(page_table.QueryInfo(std::addressof(mem_info), std::addressof(page_info),
+                                          cur_addr));
+            auto svc_mem_info = mem_info.GetSvcMemoryInfo();
+
+            // check if this memory block is heap
+            if (svc_mem_info.state == Kernel::Svc::MemoryState::Normal) {
+                if (svc_mem_info.size > region_size) {
+                    region_size = svc_mem_info.size;
+                    region_start = svc_mem_info.base_address;
+                }
+            }
+
+            // Check if we're done.
+            const uintptr_t next_address = svc_mem_info.base_address + svc_mem_info.size;
+            if (next_address <= GetInteger(cur_addr)) {
+                break;
+            }
+
+            cur_addr = next_address;
+        }
+        session.has_preallocated_area = false;
+        auto start_region = (region_size >> 15) >= 1024 ? smmu.Allocate(region_size) : 0;
+        if (start_region != 0) {
+            session.mapper = std::make_unique<HeapMapper>(region_start, start_region, region_size,
+                                                          smmu_id, impl->host1x);
+            session.has_preallocated_area = true;
+            LOG_CRITICAL(Debug, "Preallocation created!");
+        }
+    }
     return new_id;
 }
 
 void Container::CloseSession(size_t id) {
     std::scoped_lock lk(impl->session_guard);
+    auto& session = impl->sessions[id];
     auto& smmu = impl->host1x.MemoryManager();
+    if (session.has_preallocated_area) {
+        const DAddr region_start = session.mapper->GetRegionStart();
+        const size_t region_size = session.mapper->GetRegionSize();
+        session.mapper.reset();
+        smmu.Free(region_start, region_size);
+        session.has_preallocated_area = false;
+    }
+    session.is_active = false;
     smmu.UnregisterProcess(impl->sessions[id].smmu_id);
     impl->id_pool.emplace_front(id);
-    LOG_CRITICAL(Debug, "Closed Session {}", id);
 }
 
 Session* Container::GetSession(size_t id) {
diff --git a/src/core/hle/service/nvdrv/core/container.h b/src/core/hle/service/nvdrv/core/container.h
index a1fd20199..86705cbc8 100644
--- a/src/core/hle/service/nvdrv/core/container.h
+++ b/src/core/hle/service/nvdrv/core/container.h
@@ -20,6 +20,7 @@ class Host1x;
 
 namespace Service::Nvidia::NvCore {
 
+class HeapMapper;
 class NvMap;
 class SyncpointManager;
 
@@ -29,6 +30,9 @@ struct Session {
     size_t id;
     Kernel::KProcess* process;
     size_t smmu_id;
+    bool has_preallocated_area{};
+    std::unique_ptr<HeapMapper> mapper{};
+    bool is_active{};
 };
 
 class Container {
diff --git a/src/core/hle/service/nvdrv/core/heap_mapper.cpp b/src/core/hle/service/nvdrv/core/heap_mapper.cpp
new file mode 100644
index 000000000..59d993bc6
--- /dev/null
+++ b/src/core/hle/service/nvdrv/core/heap_mapper.cpp
@@ -0,0 +1,172 @@
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include <mutex>
+
+#include <boost/container/small_vector.hpp>
+#define BOOST_NO_MT
+#include <boost/pool/detail/mutex.hpp>
+#undef BOOST_NO_MT
+#include <boost/icl/interval.hpp>
+#include <boost/icl/interval_base_set.hpp>
+#include <boost/icl/interval_set.hpp>
+#include <boost/icl/split_interval_map.hpp>
+#include <boost/pool/pool.hpp>
+#include <boost/pool/pool_alloc.hpp>
+#include <boost/pool/poolfwd.hpp>
+
+#include "core/hle/service/nvdrv/core/heap_mapper.h"
+#include "video_core/host1x/host1x.h"
+
+namespace boost {
+template <typename T>
+class fast_pool_allocator<T, default_user_allocator_new_delete, details::pool::null_mutex, 4096, 0>;
+}
+
+namespace Service::Nvidia::NvCore {
+
+using IntervalCompare = std::less<DAddr>;
+using IntervalInstance = boost::icl::interval_type_default<DAddr, std::less>;
+using IntervalAllocator = boost::fast_pool_allocator<DAddr>;
+using IntervalSet = boost::icl::interval_set<DAddr>;
+using IntervalType = typename IntervalSet::interval_type;
+
+template <typename Type>
+struct counter_add_functor : public boost::icl::identity_based_inplace_combine<Type> {
+    // types
+    typedef counter_add_functor<Type> type;
+    typedef boost::icl::identity_based_inplace_combine<Type> base_type;
+
+    // public member functions
+    void operator()(Type& current, const Type& added) const {
+        current += added;
+        if (current < base_type::identity_element()) {
+            current = base_type::identity_element();
+        }
+    }
+
+    // public static functions
+    static void version(Type&){};
+};
+
+using OverlapCombine = counter_add_functor<int>;
+using OverlapSection = boost::icl::inter_section<int>;
+using OverlapCounter = boost::icl::split_interval_map<DAddr, int>;
+
+struct HeapMapper::HeapMapperInternal {
+    HeapMapperInternal(Tegra::Host1x::Host1x& host1x) : device_memory{host1x.MemoryManager()} {}
+    ~HeapMapperInternal() = default;
+
+    template <typename Func>
+    void ForEachInOverlapCounter(OverlapCounter& current_range, VAddr cpu_addr, u64 size,
+                                 Func&& func) {
+        const DAddr start_address = cpu_addr;
+        const DAddr end_address = start_address + size;
+        const IntervalType search_interval{start_address, end_address};
+        auto it = current_range.lower_bound(search_interval);
+        if (it == current_range.end()) {
+            return;
+        }
+        auto end_it = current_range.upper_bound(search_interval);
+        for (; it != end_it; it++) {
+            auto& inter = it->first;
+            DAddr inter_addr_end = inter.upper();
+            DAddr inter_addr = inter.lower();
+            if (inter_addr_end > end_address) {
+                inter_addr_end = end_address;
+            }
+            if (inter_addr < start_address) {
+                inter_addr = start_address;
+            }
+            func(inter_addr, inter_addr_end, it->second);
+        }
+    }
+
+    void RemoveEachInOverlapCounter(OverlapCounter& current_range,
+                                    const IntervalType search_interval, int subtract_value) {
+        bool any_removals = false;
+        current_range.add(std::make_pair(search_interval, subtract_value));
+        do {
+            any_removals = false;
+            auto it = current_range.lower_bound(search_interval);
+            if (it == current_range.end()) {
+                return;
+            }
+            auto end_it = current_range.upper_bound(search_interval);
+            for (; it != end_it; it++) {
+                if (it->second <= 0) {
+                    any_removals = true;
+                    current_range.erase(it);
+                    break;
+                }
+            }
+        } while (any_removals);
+    }
+
+    IntervalSet base_set;
+    OverlapCounter mapping_overlaps;
+    Tegra::MaxwellDeviceMemoryManager& device_memory;
+    std::mutex guard;
+};
+
+HeapMapper::HeapMapper(VAddr start_vaddress, DAddr start_daddress, size_t size, size_t smmu_id,
+                       Tegra::Host1x::Host1x& host1x)
+    : m_vaddress{start_vaddress}, m_daddress{start_daddress}, m_size{size}, m_smmu_id{smmu_id} {
+    m_internal = std::make_unique<HeapMapperInternal>(host1x);
+}
+
+HeapMapper::~HeapMapper() {
+    m_internal->device_memory.Unmap(m_daddress, m_size);
+}
+
+DAddr HeapMapper::Map(VAddr start, size_t size) {
+    std::scoped_lock lk(m_internal->guard);
+    m_internal->base_set.clear();
+    const IntervalType interval{start, start + size};
+    m_internal->base_set.insert(interval);
+    m_internal->ForEachInOverlapCounter(m_internal->mapping_overlaps, start, size, [this](VAddr start_addr, VAddr end_addr, int){
+        const IntervalType other{start_addr, end_addr};
+        m_internal->base_set.subtract(other);
+    });
+    if (!m_internal->base_set.empty()) {
+        auto it = m_internal->base_set.begin();
+        auto end_it = m_internal->base_set.end();
+        for (; it != end_it; it++) {
+            const VAddr inter_addr_end = it->upper();
+            const VAddr inter_addr = it->lower();
+            const size_t offset = inter_addr - m_vaddress;
+            const size_t sub_size = inter_addr_end - inter_addr;
+            m_internal->device_memory.Map(m_daddress + offset, m_vaddress + offset, sub_size, m_smmu_id);
+        }
+    }
+    m_internal->mapping_overlaps += std::make_pair(interval, 1);
+    m_internal->base_set.clear();
+    return m_daddress + (start - m_vaddress);
+}
+
+void HeapMapper::Unmap(VAddr start, size_t size) {
+    std::scoped_lock lk(m_internal->guard);
+    m_internal->base_set.clear();
+    m_internal->ForEachInOverlapCounter(m_internal->mapping_overlaps, start, size, [this](VAddr start_addr, VAddr end_addr, int value) {
+        if (value <= 1) {
+            const IntervalType other{start_addr, end_addr};
+            m_internal->base_set.insert(other);
+        }
+    });
+    if (!m_internal->base_set.empty()) {
+        auto it = m_internal->base_set.begin();
+        auto end_it = m_internal->base_set.end();
+        for (; it != end_it; it++) {
+            const VAddr inter_addr_end = it->upper();
+            const VAddr inter_addr = it->lower();
+            const size_t offset = inter_addr - m_vaddress;
+            const size_t sub_size = inter_addr_end - inter_addr;
+            m_internal->device_memory.Unmap(m_daddress + offset, sub_size);
+        }
+    }
+    const IntervalType to_remove{start, start + size};
+    m_internal->RemoveEachInOverlapCounter(m_internal->mapping_overlaps, to_remove, -1);
+    m_internal->base_set.clear();
+}
+
+} // namespace Service::Nvidia::NvCore
\ No newline at end of file
diff --git a/src/core/hle/service/nvdrv/core/heap_mapper.h b/src/core/hle/service/nvdrv/core/heap_mapper.h
new file mode 100644
index 000000000..8b23638b8
--- /dev/null
+++ b/src/core/hle/service/nvdrv/core/heap_mapper.h
@@ -0,0 +1,48 @@
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <memory>
+
+#include "common/common_types.h"
+
+namespace Tegra::Host1x {
+class Host1x;
+} // namespace Tegra::Host1x
+
+namespace Service::Nvidia::NvCore {
+
+class HeapMapper {
+public:
+    HeapMapper(VAddr start_vaddress, DAddr start_daddress, size_t size, size_t smmu_id,
+               Tegra::Host1x::Host1x& host1x);
+    ~HeapMapper();
+
+    bool IsInBounds(VAddr start, size_t size) const {
+        VAddr end = start + size;
+        return start >= m_vaddress && end <= (m_vaddress + m_size);
+    }
+
+    DAddr Map(VAddr start, size_t size);
+
+    void Unmap(VAddr start, size_t size);
+
+    DAddr GetRegionStart() const {
+        return m_daddress;
+    }
+
+    size_t GetRegionSize() const {
+        return m_size;
+    }
+
+private:
+    struct HeapMapperInternal;
+    VAddr m_vaddress;
+    DAddr m_daddress;
+    size_t m_size;
+    size_t m_smmu_id;
+    std::unique_ptr<HeapMapperInternal> m_internal;
+};
+
+} // namespace Service::Nvidia::NvCore
\ No newline at end of file
diff --git a/src/core/hle/service/nvdrv/core/nvmap.cpp b/src/core/hle/service/nvdrv/core/nvmap.cpp
index 0b2ddd980..023c070d9 100644
--- a/src/core/hle/service/nvdrv/core/nvmap.cpp
+++ b/src/core/hle/service/nvdrv/core/nvmap.cpp
@@ -8,10 +8,12 @@
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "core/hle/service/nvdrv/core/container.h"
+#include "core/hle/service/nvdrv/core/heap_mapper.h"
 #include "core/hle/service/nvdrv/core/nvmap.h"
 #include "core/memory.h"
 #include "video_core/host1x/host1x.h"
 
+
 using Core::Memory::YUZU_PAGESIZE;
 
 namespace Service::Nvidia::NvCore {
@@ -90,10 +92,19 @@ void NvMap::UnmapHandle(Handle& handle_description) {
     }
 
     // Free and unmap the handle from the SMMU
-    auto& smmu = host1x.MemoryManager();
-    smmu.Unmap(handle_description.d_address, handle_description.aligned_size);
-    smmu.Free(handle_description.d_address, static_cast<size_t>(handle_description.aligned_size));
+    const size_t map_size = handle_description.aligned_size;
+    if (!handle_description.in_heap) {
+        auto& smmu = host1x.MemoryManager();
+        smmu.Unmap(handle_description.d_address, map_size);
+        smmu.Free(handle_description.d_address, static_cast<size_t>(map_size));
+        handle_description.d_address = 0;
+        return;
+    }
+    const VAddr vaddress = handle_description.address;
+    auto* session = core.GetSession(handle_description.session_id);
+    session->mapper->Unmap(vaddress, map_size);
     handle_description.d_address = 0;
+    handle_description.in_heap = false;
 }
 
 bool NvMap::TryRemoveHandle(const Handle& handle_description) {
@@ -188,24 +199,31 @@ DAddr NvMap::PinHandle(NvMap::Handle::Id handle, size_t session_id, bool low_are
         DAddr address{};
         auto& smmu = host1x.MemoryManager();
         auto* session = core.GetSession(session_id);
-        while ((address = smmu.Allocate(handle_description->aligned_size)) == 0) {
-            // Free handles until the allocation succeeds
-            std::scoped_lock queueLock(unmap_queue_lock);
-            if (auto freeHandleDesc{unmap_queue.front()}) {
-                // Handles in the unmap queue are guaranteed not to be pinned so don't bother
-                // checking if they are before unmapping
-                std::scoped_lock freeLock(freeHandleDesc->mutex);
-                if (handle_description->d_address)
-                    UnmapHandle(*freeHandleDesc);
-            } else {
-                LOG_CRITICAL(Service_NVDRV, "Ran out of SMMU address space!");
+        const VAddr vaddress = handle_description->address;
+        const size_t map_size = handle_description->aligned_size;
+        handle_description->session_id = session_id;
+        if (session->has_preallocated_area && session->mapper->IsInBounds(vaddress, map_size)) {
+            handle_description->d_address = session->mapper->Map(vaddress, map_size);
+            handle_description->in_heap = true;
+        } else {
+            while ((address = smmu.Allocate(map_size)) == 0) {
+                // Free handles until the allocation succeeds
+                std::scoped_lock queueLock(unmap_queue_lock);
+                if (auto freeHandleDesc{unmap_queue.front()}) {
+                    // Handles in the unmap queue are guaranteed not to be pinned so don't bother
+                    // checking if they are before unmapping
+                    std::scoped_lock freeLock(freeHandleDesc->mutex);
+                    if (handle_description->d_address)
+                        UnmapHandle(*freeHandleDesc);
+                } else {
+                    LOG_CRITICAL(Service_NVDRV, "Ran out of SMMU address space!");
+                }
             }
-        }
 
-        handle_description->d_address = address;
-
-        smmu.Map(address, handle_description->address, handle_description->aligned_size,
-                 session->smmu_id);
+            handle_description->d_address = address;
+            smmu.Map(address, vaddress, map_size, session->smmu_id);
+            handle_description->in_heap = false;
+        }
     }
 
     if (low_area_pin) {
diff --git a/src/core/hle/service/nvdrv/core/nvmap.h b/src/core/hle/service/nvdrv/core/nvmap.h
index 7dd6d26c3..4af61289e 100644
--- a/src/core/hle/service/nvdrv/core/nvmap.h
+++ b/src/core/hle/service/nvdrv/core/nvmap.h
@@ -70,6 +70,8 @@ public:
 
         u8 kind{};        //!< Used for memory compression
         bool allocated{}; //!< If the handle has been allocated with `Alloc`
+        bool in_heap{};
+        size_t session_id{};
 
         DAddr d_address{}; //!< The memory location in the device's AS that this handle corresponds to,
                            //!< this can also be in the nvdrv tmem
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 6ad3b94f8..609704b33 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -34,8 +34,6 @@
 #include "video_core/renderer_base.h"
 #include "video_core/shader_notify.h"
 
-#pragma optimize("", off)
-
 namespace Tegra {
 
 struct GPU::Impl {